DEV Community: useapi.net

How to Automate NotebookLM with the Gemini Notebook API: Deep Research, Chat, Audio Overviews and a Lot More

useapi.net — Wed, 30 Sep 2026 03:43:30 +0000

Introduction

The useapi.net Gemini Notebook API is a REST API for Google's Gemini Notebook, formerly NotebookLM, that you can call from any language that can send an HTTP request. It drives your own Google account: notebooks, sources, grounded chat, notes, web research, sharing and every Studio artifact. It also runs unattended: jobs spread across your connected accounts by remaining usage and are tracked server-side, so nobody has to watch the usage limits or decide when the next batch can start (see the FAQ).

Four sources and two research questions went in, and 30 minutes of API calls later the notebook held an Audio Overview, an 11-minute Video Overview, a slide deck, an infographic, a Deep Research report and every text artifact Studio makes. Everything below is real output, and every call that made it is further down the page.

Sources and prompts

Input	What it is
Web page	Apollo 11 on Wikipedia, 27,095 words
YouTube video	A New Look at the Apollo 11 Landing Site, NASA Goddard
Pasted text	One paragraph, "Why Apollo 11 almost did not land"
PDF upload	The same Wikipedia article as a 3.8 MB PDF
Chat question	"What went wrong during the descent, and how did the crew handle it?", then "How much fuel was left when they landed?"
Discover sources	"Apollo 11 lunar module guidance computer alarms"
Deep Research	"Why did the Apollo 11 landing nearly abort?"

Results

Video Overview — explainer, whiteboard style, 11 min 50 s
the first 90 seconds, cut without re-encoding for our media host · [step 18](#18-a-video-overview-and-a-slide-deck)

Slide deck — presenter format, 8 slides, pdf + pptx
[the whole deck as PDF](https://demo.useapi.net/discord-cdn-proxy/?https://cdn.discordapp.com/attachments/1239264794394234985/1554374611972530217/gemini-notebook-tutorial-pro-slides-presenter-apollo11-deck-0929.pdf) · [step 18](#18-a-video-overview-and-a-slide-deck)

Slide 2 as generated

Slide 2 revised — "Make this slide a simple timeline graphic."
51 seconds · [step 19](#19-revise-one-slide)

Infographic — portrait, sketch note, 1536 × 2752
1 min 45 s · [step 11](#11-an-audio-overview-and-an-infographic)

🎧 Play on useapi.net

Audio Overview — brief, short, 1 min 31 s
"Fifty Seconds of Fuel and Lunar Isolation" · 4 min 36 s to generate · [step 11](#11-an-audio-overview-and-an-infographic)

And as text or JSON, ready to use in your own app:

Result	What came back	Time
Chat answer	4,237 characters with 21 numbered citations, and a follow-up in the same conversation	35 s
Discover sources	10 web pages with one-line descriptions, 3 imported	16 s
Deep Research	A 26,878-character cited report on 30 web sources, imported as a source	3 min 40 s
Quiz	6 questions with answer options, hints and rationales	33 s
Flashcards	20 cards	18 s
Mind map	A JSON tree of the whole topic	92 s
Briefing report	27,282 characters of Markdown	79 s
Data table	12 rows of alarms and anomalies with their causes	32 s
Source guide	Google's summary and key topics of the Wikipedia page	1 s

What it cost

useapi.net is an API, not a hosted Gemini Notebook. Every call runs on a Google account of your own, which you connect once through Setup Gemini Notebook, and everything it makes lives in that account. Gemini Notebook itself is free on a free Google account, and paid Google AI plans raise its limits.

You pay useapi.net a flat $15/month for API access to every supported service, and nothing per generation. Each subscription lets you connect 3 Google accounts, up to 100 in total. With more than one connected, the API load-balances across them: omit email and a new notebook goes to the least busy account, and a one-shot POST /artifacts runs on an account with a free job slot and enough usage left for the job. GET /jobs shows the load on every account.

Google meters it with a usage budget per account that refills every 5 hours, plus a weekly budget. Everything on this page used about 35% of a Google AI Pro account's 5-hour budget, across two windows (10% for the first script, 25% for the second, mostly the video and the slide deck), and 1.7% of its week. How many of each you can start with a full 5-hour budget, counting each job at Google's own estimate (so these are minimums):

Studio action	Free	Pro	Ultra
Audio Overview	2	9	181
Video Overview	2	9	183
Video Overview, short	2	11	229
Video Overview, cinematic	not available	1	25
Report	23	92	1846
Table	16	67	1355
Quiz	43	175	3478
Flashcards	52	212	4285
Infographic	9	37	745
Slide deck	1	7	146
Mind map	66	263	5217
Chat question	62	243	4897

GET /accounts/email shows each account's live usage, reset times and every action's current estimate.

How it was built

Two scripts ran back to back: the first (steps 1 to 14) built and researched the notebook and generated the audio, infographic and quiz, the second (steps 15 to 19) made everything else on the same notebook. Both are at the end of the page. Each step shows its result, and its curl request with the response it got is one click away.

You need a useapi.net API token and a Google account connected through Setup Gemini Notebook. Every call is also in the Postman collection if you'd rather click than script. The steps use these variables:

export USEAPI_TOKEN="user:12345-..."
export EMAIL="user@example.com"
API=https://api.useapi.net/v1/gemini-notebook
enc() { jq -rn --arg v "$1" '$v|@uri'; }

The ids the API returns name the account they belong to and contain : and @, so they are URL-encoded (enc) whenever they go into a URL path. Studio generation and research run as jobs: the default mode: sync waits up to 90 seconds and returns the finished result, or 202 with the running job, and mode: async returns 201 with a jobid at once. Poll GET /jobs/jobid or pass replyUrl for a webhook.

1. Create a notebook

POST /notebooks on the connected account. Omit email and the API picks the least busy of your accounts.

curl -sS -X POST "$API/notebooks" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d "$(jq -n --arg email "$EMAIL" '{email: $email, title: "Apollo 11"}')" | tee notebook.json
NOTEBOOK=$(jq -r .notebook notebook.json)

It returned, in full (step-01-notebook.json):

{
  "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
  "email": "user@example.com",
  "title": "Apollo 11"
}

2. Add a web page, a YouTube video and pasted text

One POST /sources call takes up to 50 urls (YouTube links become video sources) plus pasted text. A file in the account's Google Drive goes in the same call as drive: [{fileId, mimeType, name}], and POST /sources/sync re-imports it after it changes.

3. Upload a file

POST /sources/upload takes the file as the raw request body, with its MIME type as Content-Type: PDF, Word, PowerPoint, EPUB, Markdown, text, CSV, audio, video or an image.

curl -sS -X POST "$API/sources/upload?notebook=$(enc "$NOTEBOOK")&name=01-apollo11-wikipedia.pdf" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/pdf" \
  --data-binary @01-apollo11-wikipedia.pdf

It returned, in full (step-03-upload.json):

{
  "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
  "source": "user:12345-user@example.com-source:62600bab-4082-4ee6-8fc3-955cee0c69cf:6e04a6ea-cd2d-44cb-8be5-66f48be21a0d",
  "title": "01-apollo11-wikipedia.pdf",
  "status": "preparing"
}

4. Wait until every source is ready

Google processes each source before it can be used. The 3.8 MB PDF took about 45 seconds, the others were ready at once.

5. See what a source is about

GET /sources/source returns Google's summary and key topics of one source, and with ?content=true the full text Google indexed from it. For the Wikipedia page the topics were Apollo spaceflight, Lunar landing, Space Race, Mission personnel and Space exploration.

6. Ask a question

POST /chat answers from the sources with numbered citations, each naming its source and quoting the passage. This answer took 35 seconds and carried 21 citations. Send the same conversation with the next question to keep the context (step 15).

7. Keep the answer as a note

POST /notes saves text in the notebook. Notes can be listed, edited and deleted, and step 16 turns this one into a source.

curl -sS -X POST "$API/notes" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d "$(jq -n --arg nb "$NOTEBOOK" --arg a "$(jq -r .answer chat.json)" '{notebook: $nb, title: "What went wrong during the descent", content: $a}')" | tee note.json

It returned, in full (step-07-note.json):

{
  "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
  "note": "user:12345-user@example.com-note:62600bab-4082-4ee6-8fc3-955cee0c69cf:3ae0d70e-29eb-4f70-804e-0fe6c936f35f",
  "title": "What went wrong during the descent",
  "content": "During the lunar descent of Apollo 11's Lunar Module *Eagle*, a cascade of unexpected technical anomalies and environmental hazards threatened the landing, requiring rapid problem-solving from both the crew and Mission Control [1-6].\n\n---\n\n### **1. Overshooting the Planned Landing Target**\n* **What went wrong:** As descent began, Neil Armstrong and Buzz Aldrin noticed they were passing lunar landmarks two to three seconds early, meaning *Eagle* was traveling too fast and would land miles downrange of their intended target [3, 7]. Flight controllers suspected this was caused by lunar gravitational anomalies (mascons), extra air pressure in the docking tunnel, or unmodeled forces during earlier maneuvers [3, 7].\n* **How it was handled:** The crew monitored their position closely, maintained composure, and prepared to adapt to landing further west than planned [2, 3, 5, 7].\n\n---\n\n### **2. Computer Program Alarms (1201 and 1202)**\n* **What went wrong:** Five minutes into the descent burn, at about 6,000 feet above the surface, the Lunar Module Guidance Computer (LGC) raised unexpected **1201 and 1202 program alarms** [1, 4, 6]. These indicated \"executive overflows\"—the computer was receiving more data than it could process in real-time and had to postpone lower-priority tasks [1, 4]. The root cause was later identified as a radar hardware mismatch that fed spurious data from the rendezvous radar into the computer while the landing radar was also active [8-10].\n* **How it was handled:** \n  * **Mission Control:** Computer engineer **Jack Garman** recognized that the alarms were safe as long as they remained intermittent [1, 4]. Guidance Officer **Steve Bales** agreed, and CAPCOM **Charles Duke** relayed the \"GO\" call to the crew [1, 4].\n  * **Onboard Software:** Software engineered under **Margaret Hamilton** at MIT was designed to drop lower-priority tasks and prioritize critical guidance calculations, preventing a system crash or an automatic abort [11, 12].\n\n---\n\n### **3. Hazardous Landing Terrain (West Crater)**\n* **What went wrong:** When Armstrong looked out his window, he saw that the automated guidance system was targeting a boulder-strewn area surrounding a 300-foot-wide crater (West crater) [2, 5, 6]. Touching down there would have risked flipping or damaging the lander [2, 5].\n* **How it was handled:** Armstrong took **semi-automatic manual control** of *Eagle* to fly past the boulder field and find clear ground [2, 5, 6, 13, 14]. While Armstrong focused entirely on flying the spacecraft, **Buzz Aldrin called out altitude and navigation data** continuously to keep him informed [2, 5].\n\n---\n\n### **4. Dwindling Fuel and Blinding Lunar Dust**\n* **What went wrong:** Extending the flight to find a clear landing site consumed critical propellant [13, 14]. At 100 feet above the surface, the lander had only about **90 seconds of fuel remaining** [13, 14]. Additionally, the engine exhaust kicked up a blinding cloud of fine lunar dust that obscured the surface, making it difficult to judge motion and speed [14, 15].\n* **How it was handled:** Armstrong picked out large rocks projecting above the dust cloud to visually gauge *Eagle*'s horizontal speed [14, 15]. He cleared a smaller crater and brought the lander down smoothly with roughly **25 to 50 seconds of fuel remaining** before a forced abort would have become necessary [13, 14, 16, 17].\n\n---\n\n### **5. Engine Shutdown and Touchdown**\n* **What went wrong:** When a 67-inch probe hanging from *Eagle*'s footpads touched the Moon, Aldrin called \"Contact light!\" [10, 15, 18]. Engine shutdown was supposed to occur immediately to avoid gas pressure build-up under the craft, but in the intensity of the moment, Armstrong delayed shutting down the engine for about three seconds [15, 18].\n* **How it was handled:** *Eagle* settled safely on the Moon without damage [15, 18]. The crew immediately ran through post-landing checklists to secure the engine, after which Armstrong announced: *\"Houston, Tranquility Base here. The Eagle has landed\"* [10, 15, 19-21].\n\n💡 Would you like to explore what happened during the moonwalk itself, or look into the close calls the crew faced during their liftoff and ascent from the lunar surface?",
  "type": "note",
  "created": "2026-09-29T04:57:19.000Z",
  "updated": "2026-09-29T04:57:19.000Z"
}

8. Find more sources on the web

POST /research with type: "fast" is Gemini Notebook's Discover sources. It returned 10 web pages in 16 seconds. Nothing enters the notebook until POST /research/import, and three were picked. Two came back: Google could not fetch the third page (a Stack Exchange question), so it is left out of the answer and listed in the notebook with status error. An import waits until Google has fetched every page, about a minute and a half here. If Google takes longer, the API answers 504 after about two minutes and the import usually still completes, so check the notebook before you retry. Never retry blindly, since that adds every source twice.

9. Start a Deep Research run

type: "deep" runs Deep Research: Google reads dozens of web pages and writes a cited report. It takes minutes, so it starts async here and is collected in step 12, after the Studio jobs.

curl -sS -X POST "$API/research" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d "$(jq -n --arg nb "$NOTEBOOK" '{notebook: $nb, type: "deep", query: "Why did the Apollo 11 landing nearly abort?", mode: "async"}')" | tee deep.json
DEEP=$(jq -r .jobid deep.json)

It returned, in full (step-09-deep-research-submit.json):

{
  "jobid": "job:e2af6631-3d2b-4630-87b1-e115620e0c4e-user:12345-user@example.com-bot:gemini_notebook",
  "email": "user@example.com",
  "type": "research",
  "status": "processing",
  "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
  "created_at": "2026-09-29T04:59:08.701Z",
  "request": {
    "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
    "type": "deep",
    "query": "Why did the Apollo 11 landing nearly abort?"
  }
}

10. A quiz

POST /artifacts with type: "quiz". It finished in 33 seconds, within the 90-second sync wait, so the call returned the finished quiz: 6 questions in result.content.quiz, each with answer options, a hint and a rationale per option. Quizzes, flashcards and mind maps are JSON, ready to render in your own app.

11. An Audio Overview and an infographic

Audio and video take minutes, so both went in async and were polled. The infographic finished in 1 minute 45 seconds, the Audio Overview in 4 minutes 36 seconds. Each finished job lists its files, and every file link goes through GET /artifacts/download with your token (m4a, mp4, pdf, pptx, png). Both are shown in Results.

12. Collect the Deep Research report

The Deep Research run from step 9 finished in 3 minutes 40 seconds, while the Studio jobs ran: "Anatomy of a Lunar Descent: Systems Architecture, Hardware Anomaly, and Human Decision-Making in the Apollo 11 Landing Near-Abort", 26,878 characters of Markdown with [cite: n] markers, on 30 sources, 24 of them cited. Imported as a source (3,146 words), it grounds every later answer and artifact. urls: [] with report: true imports the report alone, and leaving urls out imports it with every cited source, as the app's Import button does.

13. Share the notebook

POST /sharing with a public link lets anyone with it view the notebook, and allowCopies: false stops them duplicating it. people: [{email, role}] invites someone as viewer or editor (Google emails them), role: "remove" takes the access away, and link: "restricted" makes the notebook private again.

curl -sS -X POST "$API/sharing" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d "$(jq -n --arg nb "$NOTEBOOK" '{notebook: $nb, link: "public", allowCopies: false}')"

It returned, in full (step-13-sharing.json):

{
  "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
  "people": [
    {
      "email": "user@example.com",
      "role": "owner",
      "name": "User"
    }
  ],
  "link": "public",
  "allowCopies": false
}

14. What the first script used

GET /accounts/email reads the account's usage. Steps 1 to 13 took the 5-hour window from 9.6% to 19.5% and its week from 7.0% to 7.5%, in 9 minutes 20 seconds end to end.

15. Ask a follow-up question

The second script starts here, on the same notebook. The conversation from step 6 makes the answer build on the earlier turns: "When the Lunar Module Eagle touched down, it had 216 pounds (98 kg) of usable fuel remaining", with 5 citations.

16. Turn the note into a source

POST /notes/source turns the note from step 7 into a Markdown source that later answers can cite.

curl -sS -X POST "$API/notes/source" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d "$(jq -n --arg n "$(jq -r .note note.json)" '{note: $n}')"

It returned, in full (step-16-note-to-source.json):

{
  "notebook": "user:12345-user@example.com-notebook:62600bab-4082-4ee6-8fc3-955cee0c69cf",
  "source": "user:12345-user@example.com-source:62600bab-4082-4ee6-8fc3-955cee0c69cf:f9aa14ef-446a-445a-8d94-b45b1b9a5045",
  "title": "What went wrong during the descent",
  "kind": "markdown",
  "status": "ready",
  "words": 656,
  "created": "2026-09-29T05:49:29.000Z"
}

17. Flashcards, a mind map, a report and a data table

Four sync calls, one after another: the flashcards in 18 seconds, the data table in 32, the report in 79 and the mind map in 92. The mind map outlived the 90-second sync wait, so it answered 202 and was polled to the end. Flashcards come back as result.content.flashcards (20 cards of front f and back b), the mind map as a JSON tree in result.content, the briefing report as 27,282 characters of Markdown in result.text, and the data table as result.table, the header row first. The table's first rows, with the cells shortened and the Source column left out here:

Time	Alarm or event	Cause	What the crew or Mission Control did
Five minutes into the descent burn / MET 102:38:00–102:38:22 / PDI + 316 seconds	1202 program alarm	Computer overload (Executive overflow: no core sets available). High-frequency counter increment requests …	Buzz Aldrin checked the alarm code using Verb 90 Noun 50 (and Verb 05 Noun 09 at PDI + 322 seconds). Computer …
MET 102:36:55	Slow spacecraft rotation rate during pitch-over/spin maneuver	Autopilot rate switch was set at 5 deg/sec.	Neil Armstrong switched the autopilot rate switch from 5 deg/sec to 25 deg/sec.
As descent began / Not in source	Overshooting the Planned Landing Target / Downrange Position Offset (Trajectory shifted 3,000 ft downrange …	Residual pressure in the docking tunnel during separation, lunar gravitational anomalies (mascons), and …	The crew monitored their position closely, maintained composure, and prepared to adapt to landing further …

18. A Video Overview and a slide deck

Both async. The slide deck finished in 5 minutes 10 seconds: 8 slides, with pdf and pptx files and one image per slide in result.slides. The Video Overview took 14 minutes and runs 11 minutes 50 seconds, a 720p mp4 of 59 MB. Both are shown in Results.

19. Revise one slide

POST /artifacts/revise redrew slide 2 from a prompt in 51 seconds. The revision is a new artifact with its own artifact id, its job's result is the whole deck again with the new slide 2, and the original deck stays in the notebook. Before and after are shown in Results.

The second script used about 25% of the 5-hour window and 1.2% of the week. The Video Overview and the slide deck are the expensive ones, at Google's estimate of about 11% and 14% each.

Shortcuts

For a single artifact you do not need to manage a notebook. Send the sources straight to POST /artifacts: the API creates a notebook on one of your accounts with a free job slot whose usage window allows the job, waits for the sources, generates, and can delete the notebook afterwards for text-only types. If it answers 202 (the job outlived the 90-second wait), poll its jobid.

curl -sS -X POST "$API/artifacts" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d '{"type": "quiz", "urls": ["https://en.wikipedia.org/wiki/Apollo_11"], "deleteNotebook": true}' \
  | jq '{status, questions: (.result.content.quiz | length)}'

To be told when a job finishes instead of polling, pass replyUrl (and optionally replyRef) with mode: async, on Studio jobs and research alike. The API sends one POST with the same job record GET /jobs/jobid returns. DELETE /jobs/jobid stops a research run you no longer need.

curl -sS -X POST "$API/research" \
  -H "Authorization: Bearer $USEAPI_TOKEN" -H "Content-Type: application/json" \
  -d "$(jq -n --arg nb "$NOTEBOOK" '{notebook: $nb, type: "deep", query: "How did the Apollo guidance computer recover from overloads?", mode: "async", replyUrl: "https://example.com/hooks/notebook", replyRef: "order-42"}')"

More of the API

The rest of the API, on an Apollo 11 notebook like the one above:

GET /notebooks/guide returned Google's summary of the whole notebook, three suggested reports such as "Explain the technical challenges faced during the Eagle's lunar descent.", and three suggested chat questions.
POST /chat/settings with the persona "You are a NASA flight controller. Answer in two sentences." made the next answers short and in that voice. GET /chat then returned the conversation's four turns, oldest first, and DELETE /chat cleared it.
POST /notes with conversation saved that whole conversation as one note, "Chat History", question and answer by turn.
POST /notes/mindmap made a mind map note, "Apollo 11 Mission", with six branches from "Background & Goals" to "Legacy & Artifacts".
POST /labels with auto: "all" grouped both sources under a "Space Exploration" label. Labels of your own take a name and an emoji, and POST /labels/label renames them and moves sources in and out.
POST /notebooks/copy copied the notebook with its two sources and two Studio artifacts.
POST /accounts with outputLanguage: "es" switched the account's Studio output to Spanish, and back.
POST /research with source: "drive" searches the account's Google Drive instead of the web. It found a Word .docx and a Google Doc about Apollo 11, and POST /research/import added them as Drive sources. After the Doc was edited in Drive, GET /sources/source reported it inSync: false, and a few seconds later the notebook held the new text. POST /sources/sync re-imports a changed Drive file on request.
A preset report takes extra instructions: a briefing "for a high-school class, in under 300 words" came back as 1,946 characters.

What the NotebookLM API can do

Step	Endpoint	What it does
Create a notebook	POST /notebooks	Creates an empty notebook on one of your connected accounts.
Add sources	POST /sources	Web pages and YouTube videos (`urls`, up to 50 per call), pasted `text` and Google Drive files (`drive`).
Upload a file	POST /sources/upload	The file is the raw request body: PDF, Word, PowerPoint, EPUB, Markdown, text, CSV, audio, video or an image.
Read a source	GET /sources/`source`	Google's summary and key topics of one source, and with `content=true` the text Google indexed.
Ask	POST /chat	An answer from the sources with numbered citations. Pass `conversation` for a follow-up.
Keep notes	POST /notes	Notes in the notebook. A note can become a source.
Research	POST /research	`fast` finds sources on the web in seconds, `deep` runs Deep Research and writes a cited report.
Import research	POST /research/import	Adds the sources a research run found, and its report, to the notebook.
Generate	POST /artifacts	`audio`, `video`, `report`, `interactive_report`, `table`, `quiz`, `flashcards`, `mindmap`, `infographic` or `slides`, each with its own options.
Revise a slide	POST /artifacts/revise	Redraws one slide of a finished deck from a prompt.
Follow a job	GET /jobs/`jobid`	The status and, once finished, the result.
Download	GET /artifacts/download	The artifact's file (`m4a`, `mp4`, `pdf`, `pptx`, `png`), streamed with your API token.
Share	POST /sharing	Invite people as viewers or editors, turn the public link on or off, and limit viewers to chat only.
Copy a notebook	POST /notebooks/copy	A copy with every source and Studio artifact, on the same account.
Notebook guide	GET /notebooks/guide	Google's summary of the whole notebook, suggested reports and suggested chat questions.
Chat history and settings	GET /chat, DELETE /chat, POST /chat/settings	A conversation's turns, delete the history, and set the chat's goal, persona and answer length.
Source labels	POST /labels	Google labels the sources by topic, or you make and edit your own labels.
Mind map note	POST /notes/mindmap	A mind map made in one call and kept as a note.
Output language	POST /accounts	The language Google writes Studio artifacts in, for the whole account.

The POST /artifacts page lists every Studio type's options.

Run both scripts yourself

Both scripts save every response under a timestamped folder. The first builds the notebook, the second takes the notebook, conversation and note ids the first one printed.

Both scripts are also in the gemini-notebook-api GitHub repo, with Node.js and Python examples that turn URLs into an Audio Overview podcast and save a Deep Research report as Markdown.

Frequently asked questions

Is Gemini Notebook the same as NotebookLM? Yes. Gemini Notebook is Google's current name for NotebookLM, and the useapi.net Gemini Notebook API works as a NotebookLM API: notebooks, sources, chat, Deep Research and every Studio artifact over REST.

Can I generate a NotebookLM Audio Overview (podcast) through an API? Yes. POST /artifacts with type: "audio" generates an Audio Overview (deep_dive, brief, critique or debate) and returns an m4a file. The short brief in this walkthrough took about 4 and a half minutes.

How does the API reach my Gemini Notebook? The useapi.net Gemini Notebook API drives your own Google account over REST, with a single useapi.net API token. You connect the account once through Setup Gemini Notebook.

Does it work on a free Google account? Yes. Every call on this page also runs on a free account, Deep Research included. Paid Google AI plans raise the usage budget and unlock Cinematic video: Google AI Pro gives an account about four times Free's budget, and Google AI Ultra about twenty times Pro's (see What it cost).

Which source types can I add? Web pages and YouTube videos by URL, pasted text, files in the account's Google Drive (Google Docs included), and uploaded PDF, Word, PowerPoint, EPUB, Markdown, text, CSV, audio, video and image files. See POST /sources and POST /sources/upload.

Why did my YouTube video end with status error? Google reads a YouTube video from its captions. A video with no captions at all, as with some official music videos, fails without a reason. Upload the audio or video file with POST /sources/upload instead, and Google transcribes it.

What is the difference between fast and deep research? fast (Discover sources) returns about ten relevant web pages with one-line descriptions in seconds. deep (Deep Research) reads dozens of pages for a few minutes and writes a cited report. Its result lists the pages it found and marks the ones the report cites. Both only suggest: nothing enters the notebook until POST /research/import.

How long does a generation take? Flashcards, a quiz and a data table take well under a minute, a mind map, a report or an infographic one to two minutes, an Audio Overview three to five minutes, a slide deck about five to seven, and Video Overviews ten minutes or more. See POST /artifacts for each type.

Do I have to watch the jobs and usage limits myself? No. Connect several Google accounts and the API spreads the work: each new job goes to an account that is healthy, has a free job slot and still has usage left for that type. Jobs are tracked on our side, so you poll GET /jobs/jobid or pass replyUrl and get the finished job posted to your webhook. When every account is out of usage, the 429 carries retryAt whenever Google names the reset time, so your code decides how to handle the load: wait until then, route the next batch to another account, or slow down. No one has to check in.

What happens when an account runs out of its usage window? Google refuses the generation. The call answers 429, or the job fails with error.code: "quota" when it started async, with the reset time when Google names it. Connect more accounts: one-shot requests skip accounts whose last-known usage blocks the job, and new notebooks go to the least busy account.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

Check our GitHub repo with code examples.

How to Generate Nano Banana 2 & Nano Banana Pro Images via the Google Flow API

useapi.net — Sun, 27 Sep 2026 17:01:07 +0000

Introduction

One synchronous POST to the useapi.net Google Flow API returns up to four Nano Banana 2, Nano Banana Pro or Nano Banana 2 Lite images, generated on your own Google Flow account. There's no Google Cloud project, no Gemini API key and no per-image metering. Image generation is included on any Google AI plan, even a free one, where the official Gemini API bills $0.034–$0.134 per image.

This guide covers the models you get, the price difference, the single-call image workflow with copy-paste curl, and a runnable Node.js script that batch-generates from a list of prompts. Want to see how the three models differ on one prompt? See the Nano Banana model comparison. Need video instead? See the sibling Veo 3.1 tutorial.

Supported models

Pick a model per request with the model field (default nano-banana-2-lite). All three models work on any Google AI subscription — including a free account.

Model id	Marketing name	References (I2I)	Aspect ratios	Notes
`nano-banana-2-lite` (default)	Nano Banana 2 Lite / Gemini 3.1 Flash-Lite Image	max 10	16:9, 4:3, 1:1, 3:4, 9:16, auto^†	Google Flow's default image model since July 2026 — stays at 16:9 in image-to-image mode unless you pass a ratio
`nano-banana-2`	Nano Banana 2 / Gemini 3.1 Flash Image	max 10	16:9, 4:3, 1:1, 3:4, 9:16, auto^†	Defaults to `auto` in image-to-image mode
`nano-banana-pro`	Nano Banana Pro / Gemini 3 Pro Image	max 10	16:9, 4:3, 1:1, 3:4, 9:16, auto^†	Highest-quality image model — defaults to `auto` in image-to-image mode

^† auto is valid only in image-to-image mode (at least one reference_* or character_* supplied) — the backend derives the ratio from the first reference image. With several references the orientation is not guaranteed, so pass an explicit ratio when every image must come back in the same shape.

All three accept count (1–4) and seed for reproducibility. If you omit model, the API uses nano-banana-2-lite, with or without references. Two deprecated aliases are still accepted: nano-banana maps to nano-banana-2, and imagen-4 maps to nano-banana-2-lite (Google removed Imagen from Flow in July 2026). Details and per-model capabilities are in the POST /images reference.

Upscaling to 2K/4K

Images generated with any of the three models can be upscaled with POST /images/upscale: pass the image's mediaGenerationId and a resolution of 2k (default) or 4k. The response is base64 in encodedImage. 2k works on any plan. 4k needs a paid Google plan, and on a free account it fails with 403 and a captcha_quality: PUBLIC_ERROR_UNUSUAL_ACTIVITY error that looks like a captcha problem but is not one, so don't retry it.

Pricing

Image generation is included. You keep your normal Google AI account — even a free one works for images (a paid plan is only needed for video and 4k upscales) — plus a flat $15/month to useapi.net for API access to every service. No per-image metering, no per-call surcharge.

Third-party Google Flow API by useapi.net vs. the official Gemini API — drive your own Google Flow subscription instead of metered, per-call API billing:

Model	Official Gemini API	useapi.net (Flow Pro)	useapi.net (Flow Ultra)
Veo 3.1 Fast — 8s clip	$0.80	~$0.40	~$0.10
Veo 3.1 Quality — 8s clip	$3.20	~$2.00	~$1.00
Veo 3.1 Lite — 8s clip	$0.40	~$0.20	~$0.05
Veo 3.1 Lite, lower priority	—	—	$0 (Ultra $199)
Omni 1.1 Flash — 8s clip	— (Flow only)	~$0.24	~$0.12
Omni 1.1 Flash — 8s clip at 360p	— (Flow only)	~$0.12	~$0.06
Nano Banana Pro — per image	$0.134	included	included
Nano Banana 2 — per image	$0.067	included	included
Nano Banana 2 Lite — per image	$0.034	included	included

Expected daily output — Ultra ($199/mo) plan. Daily averages observed on top real accounts. These are expected, not guaranteed: Google governs the underlying Flow allowances and they vary with demand.

Free workload (no Flow credits spent)	Avg generations / day*
Veo 3.1 Lite — lower priority video	~1,000
Images — Nano Banana 2 Lite, Nano Banana 2, Nano Banana Pro	up to ~500

*Averages from the busiest real accounts — not guarantees. They are set by Google's own allowances and fluctuate with demand. Applies only to the free lower-priority video queue and image generation. Credit-metered models (Veo 3.1 Fast / Quality and Omni 1.1 Flash) are not shown here — their volume is bounded by your plan's monthly Flow credits (see the credit table), not a free allowance.

Veo is billed per second on the Gemini API (an 8-second 720p clip shown) and images are priced per image. Through useapi.net you spend your own Flow plan credits and pay a flat $15/month — image generation is included on any Google AI plan, and there is no Google Cloud project, API key, or per-call metering.

Full plan and credit details are on the Google Flow overview. See the setup guide to connect your account.

How to generate a Nano Banana image with curl

You need a useapi.net API token and a connected Google Flow account. Image generation is synchronous — one POST blocks until the images are ready (typically 10–20 seconds) and returns them in the response, so there's no jobid to poll. Step two is just saving the result, from the signed URL or, when that is missing, the inline base64 image.

Submit the prompt — POST https://api.useapi.net/v1/google-flow/images:

curl -X POST "https://api.useapi.net/v1/google-flow/images" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Cat dressed like a pirate fencing with a real pirate, pirate looks absolutely terrified, this is happening on a deck of the Disney cruise ship",
    "model": "nano-banana-pro",
    "aspectRatio": "16:9",
    "count": 4,
    "seed": 123456
  }' > response.json

The call returns 200 OK with a jobId and a media array — one entry per image (count is 1–4, default 4). Each image's signed download URL is at media[].image.generatedImage.fifeUrl:

{
  "jobId": "j1731859345678i-u12345-email:jo***@gmail.com-bot:google-flow",
  "media": [
    {
      "image": {
        "generatedImage": {
          "seed": 123456,
          "mediaGenerationId": "user:12345-email:6a6f...-image:...",
          "prompt": "Cat dressed like a pirate fencing with a real pirate, pirate looks absolutely terrified, this is happening on a deck of the Disney cruise ship",
          "fifeUrl": "https://flow-content.google/image/...",
          "aspectRatio": "IMAGE_ASPECT_RATIO_LANDSCAPE"
        }
      }
    }
  ]
}

Download the image from fifeUrl (the signed URL is valid for a limited time, so download promptly). Once in a while fifeUrl is absent and the image comes back inline as base64 in encodedImage instead, so fall back to decoding that:

URL=$(jq -r '.media[0].image.generatedImage.fifeUrl // empty' response.json)
if [ -n "$URL" ]; then
  curl -o image_1.jpg "$URL"
else
  jq -r '.media[0].image.generatedImage.encodedImage' response.json | base64 -d > image_1.jpg
fi

Each generated image's mediaGenerationId can be reused as a reference in a later request — see Image references & characters below.

Want a webhook as well? Pass a replyUrl in the request body and the API also posts job status updates to that URL. The call itself still waits for the images. The payload matches the GET /jobs/jobId response, where a completed image lives at response.media[].image.generatedImage.fifeUrl.

Image references & characters

To steer a generation with an existing picture, upload it first with POST /assets/email (raw bytes, an image Content-Type, PNG/JPEG/WebP up to 20 MB). The upload response nests the reference id at mediaGenerationId.mediaGenerationId — that nested string is what you pass back as reference_1 (each of the three models accepts up to 10 references):

# 1. Upload a reference image
curl -X POST "https://api.useapi.net/v1/google-flow/assets/john%40gmail.com" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: image/jpeg" \
  --data-binary @reference.jpeg

# Response — the reference id is nested one level deep:
# {
#   "mediaGenerationId": {
#     "mediaGenerationId": "user:12345-email:6a6f...-image:ff9aa5cc-..."
#   },
#   "width": 1024, "height": 1024, "email": "jo***@gmail.com"
# }

# 2. Generate using that reference
curl -X POST "https://api.useapi.net/v1/google-flow/images" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-pro",
    "prompt": "Make this photograph look modern",
    "reference_1": "user:12345-email:6a6f...-image:ff9aa5cc-..."
  }'

For consistent identity across many images, bundle 1–2 reference images into a reusable character with POST /characters, then pass its character id as character_1…character_7. Characters mix freely with reference_* and share the same per-model image-ref budget. You can also drop @reference_1 / @character_1 markers inline in the prompt for positional grounding — each inline marker must have a matching body param.

Batch-generate with a script

Finding the right image takes many attempts, and running them by hand is tedious. The Node.js script reads a list of prompts from prompts.json, uploads any reference images, submits each request, and downloads every returned image — so you can queue a batch and come back to the winners. Because POST /images is synchronous, the script writes each result as soon as the call returns.

You need Node.js v21 or newer. Put prompts.json and google-flow-images.mjs in the same folder and run node ./google-flow-images.mjs API_TOKEN EMAIL, where API_TOKEN is your useapi.net API token and EMAIL is your connected Google Flow account email. The script looks the account up by email automatically and checks its health field before submitting.

Prefer to clone and run it locally? The complete google-flow-images.mjs and prompts.json are on GitHub in useapi/google-flow-api.

[
    {
        "prompt": "By default the nano-banana-2-lite model generates 4 image variations at 16:9."
    },
    {
        "model": "nano-banana-pro",
        "aspectRatio": "1:1",
        "count": 2,
        "seed": 123456,
        "prompt": "Nano Banana Pro (Gemini 3 Pro Image), two square variations. For all parameters see https://useapi.net/docs/api-google-flow-v1/post-google-flow-images"
    },
    {
        "model": "nano-banana-2",
        "aspectRatio": "auto",
        "reference_1": "./reference_image.jpeg",
        "prompt": "Image-to-image: pass a local reference file as reference_1 (uploaded automatically). auto aspect ratio derives from the first reference. All three models accept up to 10 references."
    },
    {
        "model": "nano-banana-2-lite",
        "aspectRatio": "9:16",
        "count": 4,
        "prompt": "Nano Banana 2 Lite portrait. Other models (nano-banana-2, nano-banana-pro) are selectable via the model field — see the POST /images docs."
    }
]

Examples

Real images generated through this same Google Flow API with Nano Banana Pro and Nano Banana 2 Lite — taken straight from our blog and model comparison.

Nano Banana Pro — text-to-image (source): "Cat dressed like a pirate fencing with a real pirate… on a Disney cruise ship."

Nano Banana Pro — 5 reference images (source): five individual portraits composited into one "K-Pop group victory celebration" group shot.

Nano Banana 2 Lite, the default model (source): the beekeeper prompt from our three-model comparison, rendered at 9:16 by the fastest of the three.

Frequently asked questions

Is there a Nano Banana 2 API? Yes — two ways. Google offers Nano Banana 2 (Gemini 3.1 Flash Image) on the official Gemini API, which meters per image on a Google Cloud project. Or use useapi.net's Google Flow API, which exposes it as nano-banana-2 and drives your own Google Flow account through a standard REST endpoint — no Cloud project, and image generation works on any Google AI plan including the free tier. See Pricing.
Is there a Nano Banana Pro API? Yes. Nano Banana Pro is Gemini 3 Pro Image, available on Google's official Gemini API, which meters per image. useapi.net exposes the same model as nano-banana-pro on the POST /images endpoint, driving your Flow account instead of metering per image — see Pricing.
Which model does the API use by default? nano-banana-2-lite (Gemini 3.1 Flash-Lite Image), with or without references. Switch per request by setting model to nano-banana-2 (Gemini 3.1 Flash Image) or nano-banana-pro (Gemini 3 Pro Image, the highest quality). All three accept up to 10 reference images and can be upscaled to 2K/4K. Nano Banana 2 Lite vs 2 vs Pro shows all three on one prompt.
Is Imagen 4 still available in Google Flow? No. Google removed Imagen from Google Flow in July 2026. The imagen-4 model id is still accepted so existing code keeps working, but those requests now generate with nano-banana-2-lite.
How much does it cost? Image generation works with any Google AI subscription or a free account, plus a flat monthly subscription to useapi.net for API access to all services. The official Gemini API meters per image, while useapi.net includes generation in the flat subscription — see Pricing above.
Do I need a captcha or a Google Cloud project? No Google Cloud project. Image generation requires reCAPTCHA, but you do not solve it yourself — the useapi.net worker solves it automatically, and your first Google Flow account comes with 300 free captcha credits as a one-time grant (powered by CapSolver). After that you configure your own provider keys via POST /accounts/captcha-providers. The script in this guide sends no captchaToken and relies on this automatic solving.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

Check our GitHub repo with code examples.

How to Generate Text-to-Speech Audio with the MiniMax (Hailuo) API

useapi.net — Fri, 25 Sep 2026 22:28:13 +0000

Three ways to get audio

Pick by how long the text is and whether you want to wait.

	Endpoint	Text limit	You get
Short and simple	POST speech/create-mp3	3,000 characters	The finished MP3 URL in one call
Anything longer	POST speech/create	5,000, or 10,000 on `turbo`	An `audioId` straight away, plus a `replyUrl` callback
Start playing sooner	POST speech/create with `"stream": true`	5,000, or 10,000 on `turbo`	The audio streamed back as it renders, so playback can start before it is finished

create-mp3 holds the connection open for the whole generation, which is why it is capped below what the models accept — 3,000 characters is about 22 seconds of waiting. For anything longer, use create.

Your first call

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "380426458095854",
    "text": "The last train leaves at midnight, and I still have not packed."
  }'

🎧 Play on useapi.net

That is speech-2.8-hd, the default, with no styling at all. Everything below is that same line with one thing changed.

Picking a voice

GET speech/voices lists what the account can use. It returns 600+ voices across 40 languages, and MiniMax keeps adding:

English, Arabic, Cantonese, Chinese (Mandarin), Dutch, French, German, Indonesian, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Turkish, Ukrainian, Vietnamese, Thai, Polish, Romanian, Greek, Czech, Finnish, Hindi, Bulgarian, Danish, Hebrew, Malay, Persian, Slovak, Swedish, Croatian, Filipino, Hungarian, Norwegian, Slovenian, Catalan, Nynorsk, Tamil, Afrikaans

Each voice carries tags for language, accent, gender and age, and tag_list filters on them. Matching is exact and case-sensitive, and a tag it does not recognise is ignored rather than rejected — so Female,Middle-Aged returns the same 310 voices as Female alone rather than none. A typo widens the result instead of narrowing it, so check the count when you add a filter. GET speech/config returns the current vocabulary.

curl "https://api.useapi.net/v1/minimax/speech/voices?is_system=true&tag_list=English" \
  -H "Authorization: Bearer …"

The voice does more to the result than any other parameter. Same text, same model, a British female voice instead of the US male above:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "236835177529411",
    "model": "speech-2.8-hd",
    "text": "The last train leaves at midnight, and I still have not packed."
  }'

🎧 Play on useapi.net

Steering the delivery

Everything from here uses a cloned voice called Scarlett, built from a one-minute sample — see Cloning a voice. The controls work identically on the built-in voices. Each sample has the exact call that produced it folded up above it.

Three kinds of markup go directly inside text, and they combine. Only emotion tags are free, because they are stripped before billing. Sounds and pauses are billed in full, markup and all.

Emotion

Wrap a span in {emotion}…{/emotion}. Eight are available — happy, sad, angry, fearful, disgusted, surprised, neutral, fluent — and a single request may carry up to 15 spans.

The five below are the ones that separate most cleanly. surprised and happy tend to converge, as do neutral and sad, so pick across the range rather than adjacent to it. The line is the same six words every time — only the delivery decides whether it is an accusation, dread, revulsion, delight or grief.

{angry} — accusation:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "{angry}Look at what you have done.{/angry}"
  }'

🎧 Play on useapi.net

{fearful} — dread:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "{fearful}Look at what you have done.{/fearful}"
  }'

🎧 Play on useapi.net

{disgusted} — revulsion:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "{disgusted}Look at what you have done.{/disgusted}"
  }'

🎧 Play on useapi.net

{happy} — delight:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "{happy}Look at what you have done.{/happy}"
  }'

🎧 Play on useapi.net

{sad} — grief:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "{sad}Look at what you have done.{/sad}"
  }'

🎧 Play on useapi.net

How far a voice can swing is set by the voice, not the tag. A clone inherits the range of whatever you fed it, so a one-minute sample of even-toned speech will not shout. The built-in voices tagged as expressive have more headroom — the same five emotions on Radiant Girl (236835177529409), for comparison: angry, fearful, disgusted, happy, sad.

Sounds

Nineteen non-speech sounds render in place when written in parentheses: laughs, chuckle, coughs, clear-throat, groans, breath, pant, inhale, exhale, gasps, sniffs, sighs, snorts, burps, lip-smacking, humming, hissing, emm, sneezes.

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "I told you this would happen (sighs). Every single time (laughs)."
  }'

🎧 Play on useapi.net

Pauses

<#seconds#> inserts a silence, anywhere from 0.01 to 99.99 seconds. This is the one that most improves long narration — models tend to run sentences together, and a beat between them fixes it.

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "I told you this would happen. <#1.5#> Every single time."
  }'

🎧 Play on useapi.net

Speed and pitch

speed runs 0.5 to 2 and pitch runs -12 to 12, both independent of the markup above. Back to the longer line so the difference is easier to hear.

"speed": 0.7 — 7.6 seconds:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "The last train leaves at midnight, and I still have not packed.",
    "speed": 0.7
  }'

🎧 Play on useapi.net

"speed": 1.5 — 3.2 seconds:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "The last train leaves at midnight, and I still have not packed.",
    "speed": 1.5
  }'

🎧 Play on useapi.net

"pitch": -6:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "The last train leaves at midnight, and I still have not packed.",
    "pitch": -6
  }'

🎧 Play on useapi.net

"pitch": 6:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "The last train leaves at midnight, and I still have not packed.",
    "pitch": 6
  }'

🎧 Play on useapi.net

There are also voice effects — deepen_lighten, stronger_softer and nasal_crisp take -100 to 100, and spacious_echo, lofi_telephone, robotic and auditorium_echo are switches.

Choosing a model

Model	Max characters	Credits per character	Status
`speech-2.8-hd`	5,000	1	Current, and the default
`speech-2.8-turbo`	10,000	0.6	Current
`speech-2.6-hd`	5,000	1	Previous generation
`speech-2.6-turbo`	10,000	0.6	Previous generation
`speech-2.5-hd-preview`	5,000	1	Superseded by 2.6
`speech-2.5-turbo-preview`	10,000	0.6	Superseded by 2.6
`speech-02-hd`	5,000	1	Legacy
`speech-02-turbo`	10,000	0.6	Legacy
`speech-01-hd`	5,000	1	Legacy
`speech-01-turbo`	10,000	0.6	Legacy

Use speech-2.8-hd unless you have a reason not to. The older models stay selectable because they are still live at MiniMax, and because regenerating audio to match something made earlier needs the model that made it — not because they are worth picking for new work.

Within a generation, hd and turbo are the same model tuned differently. turbo costs 40% less per character and accepts twice the text in a single call. It is not measurably faster — on identical 990-character text the two came out level — so pick turbo for the cheaper rate and the higher ceiling, and hd when voice quality matters most.

Sending more than a model accepts returns 400 before anything is generated, so an oversized request never costs credits. The message names the model and its limit — Parameter text length (5001) exceeds 5000 characters for model speech-2.8-hd — except at the absolute 10,000 ceiling, which is checked first and reports the limit without the model.

hd and turbo are the same family tuned differently. Turbo on the same line:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-turbo",
    "text": "The last train leaves at midnight, and I still have not packed."
  }'

🎧 Play on useapi.net

For narration where nobody is waiting, speech-2.8-turbo is usually the better trade — it takes twice the text per call and costs 0.6 credits a character against hd's 1, which MiniMax mirrors as $60 against $100 per million on their own API.

Word timings for captions

A finished generation carries a subtitles array with millisecond boundaries per phrase.

"subtitles": [
  { "text": "The last train leaves at midnight", "time_begin": 0, "time_end": 2245.2 }
]

Use time_begin and time_end. The text_begin and text_end fields come back from MiniMax but do not reliably index into the text, so treat them as opaque.

One limitation worth planning around: timings only reach you through POST speech/create-mp3 or the streaming done event. MiniMax does not store them, so polling GET speech/audioId or waiting on a replyUrl callback gives you has_srt but not the timings themselves. If you need captions for more than 3,000 characters, use stream and keep the done payload.

Long text without waiting

POST speech/create answers in about four seconds with an audioId, whatever the length, and keeps generating in the background.

curl "https://api.useapi.net/v1/minimax/speech/create" \
  -H "Authorization: Bearer …" \
  -H "Content-Type: application/json" \
  -d '{
    "voice_id": "380426458095854",
    "model": "speech-2.8-turbo",
    "text": "…up to 10,000 characters…",
    "replyUrl": "https://your.app/webhook"
  }'

{ "audioId": "user:12345-minimax:1234…-audio:178737039436895391", "statusLabel": "pending", "statusFinal": false }

Then either wait for the POST to replyUrl, or poll GET speech/audioId until statusFinal is true. A 10,000-character request takes around 78 seconds and yields 12 to 16 minutes of speech, so poll every few seconds rather than tightly.

Streaming

Set stream to true and the response is text/event-stream instead of JSON, with the id in an X-Audio-Id header. First audio lands about four seconds in rather than at the end.

The audio events carry two different things, and format tells them apart. format: "mp3" events are chunks — append them in arrival order and the result is a plain MP3. A short line sends one, a 10,000-character generation sends around 40 over a minute or two. format: "wav" with complete: true is the same speech again as a whole WAV file, delivered after the chunks. Do not append that one.

const chunks = [];

if (event === "audio" && payload.format === "mp3")
  chunks.push(Uint8Array.from(atob(payload.chunk), c => c.charCodeAt(0)));

A streamed generation is tracked like any other, so if the connection drops the recording still completes and the audioId still resolves. Full worked example on the endpoint page.

Cloning a voice

Every sample from Steering the delivery onward uses a cloned voice. This is how it was made.

Two calls. POST speech/upload-sample takes the raw bytes of an MP3 or WAV and returns a fileId, then POST speech/clone-voice turns that into a voice you pass as voice_id anywhere above.

curl "https://api.useapi.net/v1/minimax/speech/upload-sample" \
  -H "Authorization: Bearer …" \
  -H "Content-Type: audio/mpeg" \
  --data-binary "@sample.mp3"

{ "fileID": "user:12345-minimax:1234…-file:54768197611603968682433" }

Feed that straight into the clone:

curl "https://api.useapi.net/v1/minimax/speech/clone-voice" \
  -H "Authorization: Bearer …" \
  -H "Content-Type: application/json" \
  -d '{
    "voice_name": "Scarlett",
    "language_tag": "English",
    "files": "user:12345-minimax:1234…-file:54768197611603968682433"
  }'

{ "voice_id": "user:12345-minimax:1234…-voice:433688037470324", "voice_name": "Scarlett", "voice_status": 2 }

This is the source sample that was uploaded — one minute of ordinary speech, nothing special:

🎧 Play on useapi.net

And this is the clone reading a line it has never seen, 35 seconds after that upload:

curl 'https://api.useapi.net/v1/minimax/speech/create-mp3' \
  --header 'Authorization: Bearer user:1234-…' \
  --header 'Content-Type: application/json' \
  --data '{
    "voice_id": "user:12345-minimax:1234…-voice:433688037470324",
    "model": "speech-2.8-hd",
    "text": "The last train leaves at midnight, and I still have not packed."
  }'

🎧 Play on useapi.net

The returned voice_id goes anywhere voice_id is accepted, exactly like a built-in one.

POST speech/upload-sample takes ?account= in the query string, like the read endpoints — there is no identifier to infer it from, since the body is raw bytes. The fileID it returns then pins the voice to that account, which is why POST speech/clone-voice takes no account of its own and reads it from files. POST speech/delete-voice works the same way, taking only a voice_id.

Slots

Cloned voices occupy a slot, and how many you get depends on the MiniMax plan. Check before you spend 35 seconds discovering there is no room:

curl "https://api.useapi.net/v1/minimax/speech/equity" -H "Authorization: Bearer …"

{ "used": 2, "total": 3, "voice_clone_preview_free_quota": "inf" }

Free one with POST speech/delete-voice. Note used is allowed to exceed total — treat these as advisory and let the clone call itself be the authority.

Only clone voices you have permission to use.

Managing what you have generated

GET speech lists every recording on the account, newest first, including any made on minimax.io directly. Each row carries the audio_url, the voice name, and has_srt / has_wav telling you whether timings and a WAV rendition exist.

curl "https://api.useapi.net/v1/minimax/speech/?page_size=10" \
  -H "Authorization: Bearer …"

GET speech/audioId returns one recording in full — the voice it used, the model, cost_credit, and the speed, pitch and effects it was generated with. It is also what you poll after an async create.

DELETE speech/audioId removes one permanently. A generation still in flight cannot be deleted, because there is nothing stored yet — cancel that with DELETE scheduler/id instead.

What it costs

Speech bills per character against the MiniMax Audio balance — the same pool music draws on. One credit per character on an hd model, 0.6 on a turbo one.

That works out to $33 a million characters on the $99 Pro plan, against $100 on MiniMax's own pay-as-you-go API — the comparison at the top of this page has all three options side by side.

A few things worth knowing:

cost_credit appears only on GET speech/audioId, not on the list or on the create-mp3 response — that one returns usage_characters, the raw count before the model's ratio.
Emotion tags are free. Sound and pause markup is billed as written: 79 characters including (sighs) and (laughs) billed as 79.
turbo is charged at 0.6 credits a character, and the discount is real credits rather than a list price. GET speech/audioId reports what was charged as cost_credit.
Speech needs a MiniMax audio subscription. Without one the call returns 412 however much video credit the account holds, though a new account gets a small allowance first.
Only a few generations run concurrently per account. Past that a request returns 429 and the ones already running are unaffected, so retry when one finishes. The ceiling looks plan-dependent — a free account allows three.

Examples

Every recording on this page is a real generation, and the call that produced it is folded up directly above it. The first two use built-in voices. Everything from Steering the delivery onward uses the cloned Scarlett voice on speech-2.8-hd unless noted, apart from the long-text example, which is back on a built-in voice.

Clone source and result: the one-minute sample that was uploaded, and the clone reading a line it had never seen.

Emotion, the same six words each time — angry, fearful, disgusted, happy, sad. The same five on the built-in Radiant Girl, which has more range: angry, fearful, disgusted, happy, sad.

Everything else: sounds, pauses, slow, fast, low pitch, high pitch, turbo.

Where to go next

POST speech/create — every parameter, response and error
GET speech/voices — the voice catalogue
GET speech/config — emotion and tag vocabularies, for building your own picker
minimax-api GitHub repo — a runnable batch script that reads a list of lines, generates each one, and downloads the MP3s, with a --voices mode for finding a voice_id

Sora 2 API Alternatives Compared, With Code: Seedance 2.5, Omni Flash, MiniMax H3, Kling v3

useapi.net — Mon, 21 Sep 2026 03:48:43 +0000

Introduction

OpenAI removes sora-2, sora-2-pro and the whole Videos API on September 24, 2026, and its deprecation table names no replacement — so the migration is to another vendor, and through the useapi.net APIs the same six-second shot costs between $0.08 and $1.80 depending on which one you pick.

The most expensive of the four costs 22 times the cheapest — $1.80 against $0.08 — from the same brief, the same character and the same framing. At any real volume that gap, not the output quality, is what decides the choice.

So the comparison is not just a table, the same scene was shot five times with one character: a neon-sign maker retires a SORA 2 sign, then presents each replacement in turn. Every clip was generated by the model whose name is on the sign, so what you are watching is each model describing itself, at its own price:

The result — 31 seconds, five models, one character
$3.68 in total, all five clips · [watch on YouTube](https://www.youtube.com/watch?v=W5vXcM2cCJs)

What is actually being removed

On March 24, 2026 OpenAI notified developers that the Videos API and every Sora 2 alias and snapshot — sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08, sora-2-pro-2025-10-06 — would be removed on September 24, 2026. The Sora consumer app had already closed on April 26, 2026. After the removal date the endpoints stop answering, and OpenAI says it will permanently delete the data associated with your use of Sora once any final export window closes — so anything you still need has to come down before then.

The detail that decides the migration is the shape of the deprecations table itself. Most rows name a recommended replacement. The Sora 2 and Videos API rows leave that column empty, because OpenAI has no other video model to move you to. There is no in-house upgrade path — the only question is which vendor you go to instead.

The four replacements

Each of these generates video from a start frame with synchronised speech, which is the capability most Sora 2 pipelines were using. Prices are for a six-second clip at each model's nearest tier to 720p, on the best per-credit rate of each route.

Model	Vendor	Route	Six seconds	Notes
Omni 1.1 Flash	Google	Google Flow API	$0.08	generates at 360p or 720p, free upscale to 1080p
Kling v3	Kuaishou	Kling API	$0.44	up to 4K, derives aspect from the frame
MiniMax H3	MiniMax	PixVerse API	$0.72	768p or 1440p, no 720p tier
Seedance 2.5	ByteDance	PixVerse API	$1.80	reference audio and video, via create-fusion

MiniMax H3 is also the only one of the four whose weights are public, which matters if you are thinking about self-hosting. The FAQ covers what the licence actually allows.

Pricing

Every route here is flat-fee. You bring your own vendor account and add a flat $15/month to useapi.net for API access, with no per-generation surcharge from us — the credits above are the vendor's own, spent on your own plan.

The figures use the best per-credit rate on each route, on monthly billing: PixVerse Premium at $0.00400 per credit, Google AI Ultra $199 at roughly $0.008 per Flow credit, and Kling Ultra monthly at $0.00615. These are the lowest per-credit rates, not the cheapest plans — Premium is $60 a month and Ultra $199 is Google's top tier. Yearly billing is cheaper again on all three. Full per-model, per-plan tables are on the PixVerse overview, the Google Flow overview and the Kling overview, and the numbers move when vendors change their promotions, so treat them as current-as-of-publication rather than contractual.

Route matters as much as plan. MiniMax H3 also runs on our own MiniMax API at 7 credits per second, about $0.32 for the same six seconds against $0.72 on PixVerse. Both PixVerse clips here were kept on one account for convenience, not because it was the cheapest way to buy them. The model matrix lists every API that carries a given model.

Getting the video back

Every create call below is asynchronous — it returns an id, not a file. All of them authenticate the same way, with your useapi.net token as Authorization: Bearer user:1234-… (see Setup), and that one token reaches all four vendors.

Route	Poll	Finished when	The file
PixVerse	GET /videos/{video_id}	`video_status_final: true`	`url`
Google Flow	GET /jobs/{jobId}	`status: "completed"`	`fifeUrl` on the returned media
Kling	GET /tasks/{task_id}	`status_final: true`	`works[0].resource.resource`

Any create call also accepts replyUrl, which calls you back when the job finishes instead of making you poll.

Step 1: one character, reused everywhere

Everything starts from a single generated portrait. Images are free on Google Flow — Nano Banana Pro is billed at $0.134 an image on the official Gemini API and included here — so this stage costs nothing however many times you run it. It is passed as reference_1 to every later image request, which is what keeps the same face, the same room and the same bench across five clips made by four different vendors. Nothing else carries identity between them.

The character is invented rather than photographed. In our runs Google's moderation was markedly stricter with real, identifiable people, and a generated reference sidesteps that entirely.

The character — Nano Banana Pro
0 credits

Step 2: a start frame per model

One image request per clip, each passing the character as reference_1 and changing only the sign's text and colour. Holding the wording identical — "Same man, same room, same framing as the reference image … change only the sign" — is what keeps the framing stable enough that the five clips cut together.

Text on a generated sign is the fragile part. Twelve characters (SEEDANCE 2.5) rendered cleanly here, but that is close to where image models start mangling glyphs, so check every frame before spending a credit on video. Pick the variant where the whole sign sits inside the frame, not one that runs off the edge — once the clip animates, anything near the border clips out of shot.

Seedance 2.5

Omni 1.1 Flash

MiniMax H3

Kling v3

The start frames live in Google Flow, and the video calls below are on three different vendors. Omni 1.1 Flash takes the Flow mediaGenerationId directly. For the others, download the frame with GET /assets/{mediaGenerationId}, then upload the file to POST /files for PixVerse and use the returned path, or to POST /assets for Kling and use the returned URL.

Step 3: the farewell clip, on Sora 2 itself

The opening clip was generated on Sora 2 while it still answers, through PixVerse rather than OpenAI directly. The sign maker boxes the sign and ships it off. Sora 2 only accepts 4, 8 or 12 seconds — there is no six-second option — so this one is eight, trimmed back to six in the reel.

Sora 2 — 720p
$0.64 for eight seconds

Step 4: one clip per replacement

The same prompt shape drives all four, with only the model name and the closing line changed. Each names itself, gives a flat endorsement, then an honest catch.

Seedance 2.5
$1.80 Omni 1.1 Flash
$0.08 MiniMax H3
$0.72 Kling v3
$0.44

What changes when you move off Sora 2

Duration grids differ, and a hard-coded value will not port. Sora 2 took 4, 8 or 12 seconds. Omni 1.1 Flash takes 4, 6, 8 or 10, MiniMax H3 5 to 15, Kling v3 3 to 15, and Seedance 2.5 anything from 4 to 30.
Resolution tiers are not uniform either. Omni 1.1 Flash generates at 360p or 720p, MiniMax H3 skips 720p entirely and jumps 768p to 1440p, Seedance 2.5 reaches 1080p and Kling v3 goes to 4K. Mixing them means conforming everything to the lowest common size, so pick that size before you shoot.
Quality varies more than the resolution numbers suggest. At the same nominal size the H3 clip came back at 1.3 Mbps against Seedance's 6.5 Mbps, and it is visibly the softest of the set.
Voice is the one capability that does not port cleanly. Seedance 2.5 is the only one of the four that accepts an uploaded audio clip as a reference, and only through create-fusion. Omni 1.1 Flash offers Google's preset voices through referenceAudio_1, and Kling can carry a voice on a character element. Each clip here uses its model's own native voice, which is why they differ.
Prompts mostly carry over, with one habit worth adopting: give each physical action its own sentence and describe expressions positively. Trailing clauses and negations get dropped, unevenly, by all four.

What it cost

Step	Model	Credits	Cost
Character portrait, 9:16	Nano Banana Pro	0	$0
Four start frames, 9:16	Nano Banana Pro	0	$0
Clip 1, Sora 2, 8s 720p, trimmed to 6s	`sora-2`	160	$0.64
Clip 2, Seedance 2.5, 6s 720p	`seedance-2.5`	450	$1.80
Clip 3, Omni 1.1 Flash, 6s 720p	`omni-flash`	10	$0.08
Clip 4, MiniMax H3, 6s 768p	`minimax-h3`	180	$0.72
Clip 5, Kling v3, 6s 1080p	`kling-v3-0`	72	$0.44
Total			$3.68

Images cost nothing on Google Flow, so iterate on the character and the signs as long as you like before spending anything on video. All five clips together came to $3.68, and the cheapest of them is the $0.08 one.

Run it yourself

Every call on this page chains into two scripts. useapi/sora-2-alternatives holds them along with every prompt above in prompts.json: frames/ generates the character and a start frame per model, and clips/ sends each frame to the model named on its sign, then prints the credit cost the API returned.

The frame stage is free, so you can iterate on the character and the signs as long as you like before spending anything. One command does the cheap end of the comparison on its own:

node clips.mjs <API_TOKEN> omni

Frequently asked questions

When exactly does the Sora 2 API stop working? September 24, 2026. OpenAI announced it on March 24, 2026, covering the Videos API and every sora-2 alias and snapshot. After that date the endpoints stop answering and OpenAI permanently deletes the associated data once any final export window closes, so download anything you still need first.
What does OpenAI recommend migrating to? Nothing. The deprecations table leaves the recommended-replacement column empty for both rows, because OpenAI has no other video model in the API. Migration means another vendor.
Which alternative is closest to Sora 2 in capability? For the common case — a start frame plus synchronised speech — all four here do it. Seedance 2.5 is the most capable, taking up to 30 reference images plus 10 reference videos and 10 reference audios — 50 in total, through create-fusion — and it is also the most expensive at $1.80 per six seconds. Omni 1.1 Flash does the same core job for $0.08 with a 720p ceiling.
Is MiniMax H3 really open source? Its weights are public — MiniMax published them in early August 2026, 33B parameters, on Hugging Face. It is open-weight under a custom community licence whose territory clause requires an application form for the USA, EU, UK and South Korea, so "open source" is a loose description of it. The full repository runs to hundreds of gigabytes, though the smallest usable checkpoint is far smaller.
Do I pay useapi.net per generation? No. You bring your own vendor account and pay a flat $15/month for API access. The credits in the table are the vendor's own, spent against your plan. See Pricing.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

Check our GitHub repo with code examples.

How to Make a UGC Product Video with the Google Flow API

useapi.net — Sat, 05 Sep 2026 04:53:54 +0000

Introduction

A UGC product video can be built end to end through the useapi.net Google Flow API for 45 Flow credits, with nothing processed locally. One invented presenter shows three invented products to the camera.

UGC is the user-generated-content style that fills social feeds and ad libraries: one person talking to a phone camera, to sell a product, to entertain, or just for fun.

This is the full A-to-Z guide to producing that kind of video at a professional level: the products, the presenter, her lines and the final cut all under your control, in six steps of API calls you can copy as they are.

Here is the result, three 10-second Omni 1.1 Flash clips upscaled to 1080p, every frame and every sound produced through the Google Flow API. The same upscale call goes to 4K when you need professional quality, see Step 5:

The result — 29 seconds, 1080p, three products, three clips
45 credits in total

If you would rather watch than read, here is the whole guide in 75 seconds:

The 75-second walkthrough · watch on YouTube · narration and music from the MiniMax API

Pricing

Omni 1.1 Flash video needs a paid Google AI plan plus a flat $15/month to useapi.net for API access, with no per-call surcharge. Every image, portrait, character and the concatenate below cost nothing. The three 10-second clips cost 15 Flow credits each, 45 in total, which is 45 cents on the Google AI Ultra $99 plan. Full per-tier credit costs are on the Google Flow overview.

About the ids in the requests

Every request below refers to earlier results by their mediaGenerationId. The ids are long, so they are shortened here to user:1234-email:…-image:b0bdf3e8-…, and a line under each request says which earlier result each one is. An image id comes back from POST /images under media[0].image.generatedImage.mediaGenerationId, a character id in the character field of POST /characters, a video id under media[0].mediaGenerationId of the finished job, and an upscaled clip's id is its source's id with _upsampled on the end.

Step 1: a sheet for each product

A product sheet is one image that shows the item from four sides, so the model has the whole object and not one lucky angle. POST /images, square, one request per product. The prompt spells out the camera for every panel, because "front, side, back" on its own gave two identical panels for the mug and no true side for the keyboard. A flat object like the keyboard needs a top-down panel, a cube-shaped one does not. No hands, no captions, no logos, so nothing but the product can leak into the clips.

The same works for a real product. Upload photos of it with POST /assets and pass their ids as reference_1 to reference_4 on this request, and the model turns them into a clean four-panel sheet. The same trick makes a sheet for a package, a logo or a set.

The mug — Nano Banana 2 Lite, id b0bdf3e8
0 credits

The desk robot — Nano Banana 2 Lite, id 7aaed056
0 credits

The keyboard — Nano Banana 2 Lite, id 4f861c23
0 credits

Images cost nothing, so ask for count of 2 to 4 and keep the best sheet. The keyboard took a few rounds before its prompt described a real macro pad with a top-down panel, the cube worked first time.

Step 2: the presenter

Three calls. A front portrait from a description. A three-quarter portrait generated from the first one with reference_1, so the two agree on the face. Then POST /characters with both portraits, which returns a character reference. From then on she is @character_1 in any prompt, and Google pins her through its own entity system instead of through a description.

Front — Nano Banana 2 Lite, id a236ef66
0 credits

Rendered a few looks, picked this one. If a detail needs changing afterwards, the chain say, re-render her with her own portrait as reference_1 and a prompt that names only the change, and the face stays put.

reference_1 is the front portrait, a236ef66. This prompt still mentions freckles, a sweater and earrings left over from an earlier presenter, and it did no harm: with a reference image in the request, the reference wins over the words.

Three-quarter — id d891301c
0 credits

imageReference_1 is the front portrait a236ef66, imageReference_2 the three-quarter d891301c. The response's character field is her id, 53ca1036 below. The voice is one of Google's system presets and rides along with the character, but it cannot reach the clip mode used in Step 4, see What to expect.

Step 3: the still, the whole scene in one frame

One request puts everything together. The character goes in as character_1, the three sheets as reference_1 to reference_3, and the prompt places them by tag, left to right on the table in front of her. This frame becomes both ends of every clip, so it is the one place where the tags do their work.

character_1 is the character from Step 2, 53ca1036. reference_1 to reference_3 are the three sheets from Step 1: the mug b0bdf3e8, the desk robot 7aaed056, the keyboard 4f861c23.

The still — her face from the character, each product from its sheet, id 034cddc8
0 credits

The mug has its glow, the robot its face and antenna, the keyboard its three colours and its port, and she is the woman from the portraits.

Step 4: one clip per product, all from the still

Omni's first-plus-last-frame mode with the still in both slots, one request per product, the same POST /videos each time. This mode takes nothing but the two images, no sheets, no character, no voice reference. That is fine, because the still already holds all of it. Every item she picks up came out matching its sheet from the still alone.

The prompt does three jobs. It pins both ends to the frame, with a quiet beat on each side. It gives her the line and the voice. And it times the choreography: she starts talking as she lifts the item, the item is back on its spot by the eight-second mark, and the last two seconds are a still, silent hold. Without that last part she ran late and her arm snapped into the end frame.

startImage and endImage are both the still from Step 3, 034cddc8. The same two ids go into all three clips.

Clip 1, the mug — 10 seconds, id 55808825, shown after the 1080p upscale of Step 5
15 credits Clip 2, the desk robot — 10 seconds, id c4d0e6b1, shown after the 1080p upscale of Step 5
15 credits Clip 3, the keyboard — 10 seconds, id 3cc8ea2f, shown after the 1080p upscale of Step 5
15 credits

Same woman, same table, same light, and each clip opens and closes on the still, so any of them can follow any other. The lines are around 20 words, which is what 10 seconds holds at this pace with room for the beat at the end.

Step 5: upscale to 1080p

Omni 1.1 Flash renders at 720p. POST /videos/upscale takes a clip to 1080p for free on any paid plan, in about 40 seconds, and answers in the same call with the upscaled clip's id and URL. One request per clip, no email, the account comes from the id.

mediaGenerationId is the mug clip from Step 4, 55808825. The same call with c4d0e6b1 and 3cc8ea2f upscales the robot and the keyboard. Each response carries the new clip under media[0].mediaGenerationId, which is the source id with _upsampled on the end, so the three upscaled ids are 55808825-…_upsampled, c4d0e6b1-…_upsampled and 3cc8ea2f-…_upsampled. The clips shown in Step 4 are these upscaled versions. 4K is also available, for 50 credits and only on a Google AI Ultra plan.

Step 6: join them

POST /videos/concatenate takes the three upscaled ids in order and returns the joined file as base64, at the clips' resolution. The trims decide whether the joins feel like one take. Each clip opens with a short hold on the still and closes with a settle back onto it, and untrimmed those stack into a freeze at every cut. But she finishes each line with her hands already resting, so a trim measured from the picture alone cuts into the last words. Cut after the last sound plus a beat, never earlier. On these clips that meant leaving clip 1 whole, since its laugh runs to the last frame, taking just over half a second off the end of clip 2, and a few frames off the start of clips 2 and 3.

The three ids are the upscaled clips from Step 5 in order: the mug, the desk robot, the keyboard.

Joined — 28.8 seconds, 1080p
0 credits

She finishes a line with her hands on the table, the cut lands, and she reaches for the next item. The joined file was re-encoded to a smaller size for hosting, the three clips are as the API returned them.

What to expect

Generate, do not edit. The UGC talking head post extended a clip with video-to-video edits, and that still works when the lines are short. Here it did not. An edit keeps its source's speech on the source's own timeline, so with a 10-second line filling the seed, a new line got braided into the old one instead of replacing it, and the products left on the table drifted between takes. Generating each clip from the still gave complete lines, intact products and cleaner joins, for fewer credits.
The still carries the tags. The first-plus-last-frame mode accepts only the two images. Do the tagging in the still, with the character and the sheets, and the clips inherit it.
Time the choreography. A 10-second clip with a full line needs the put-down called out by the clock, "by the eight-second mark", and a still, silent hold at the end. Left to itself the model ran the line long and jumped into the end frame.
Describe the voice the same way in every prompt. Nothing carries a voice between separate generations, and a character can hold a voice preset but cannot reach this mode. With one identical description the three clips came back close enough to pass as one take.
Sheets need per-panel cameras. Name what each of the four panels shows. Give flat objects a top-down panel. Keep hands, captions and logos out, because whatever is on the sheet can end up in the clip.
Trim to the sound, not the picture. She goes still before she stops talking. A cut placed at the first still frame took the end of a line with it.
The presenter is invented. Google's moderation is stricter with real, identifiable people. Start from a generated character.

What it cost

Step	Mode	Credits
Three product sheets, 1:1	Nano Banana 2 Lite	0
Two presenter portraits, 3:4	Nano Banana 2 Lite	0
The character	—	0
The still, 9:16	Nano Banana 2 Lite, character + 3 references	0
Clip 1, the mug, 10s	I2V, first + last frame	15
Clip 2, the desk robot, 10s	I2V, first + last frame	15
Clip 3, the keyboard, 10s	I2V, first + last frame	15
Upscale, three clips to 1080p	—	0
Concatenate	—	0
Total		45

A 10-second Omni clip costs 15 credits at 720p, so the whole video is 45 cents on the Google AI Ultra $99 plan. Images, characters, the 1080p upscale and the concatenate cost nothing, so iterate freely on the sheets, the presenter and the still before spending a credit on video. 4K is the one upgrade that costs, 50 credits a clip on Ultra.

Run it yourself

Every step above is one API call, and the six of them chain into a script. The complete Node.js pipeline, a Python port, and the prompts.json holding every prompt on this page live in useapi/google-flow-api/ugc-product-video. Clone it, put your own products and lines in prompts.json, and one command builds the whole video.

It checkpoints after every step, so an interrupted run resumes instead of paying for the clips again, and you can redo a single step — a better still, a different presenter — without regenerating what came before it.

Frequently asked questions

Can I use my own product photos instead of generated sheets? Yes. Upload them with POST /assets and pass the returned ids as reference_1 to reference_3 on the still, exactly as the generated sheets are passed here. The four-panel sheet is a way to give the model every side of an object, and a real product benefits from the same treatment.
Why not edit one clip into the next, like the talking-head post? Because a video-to-video edit keeps its source's speech on the source's timeline. With short lines that replacement works. With a 10-second line filling the clip, the new line got braided into the old one and the products on the table drifted. Generating every clip from the same still avoids both, and costs 15 credits instead of 20. See What to expect.
Why do all three clips start and end on the same frame? So they can be joined in any order with no visible cut. The first-plus-last-frame mode pins both ends of each clip to the still, and the concatenate trims the short holds around each join. The mechanism is explained in the UGC talking head post.
Does the voice stay the same across the clips? Nothing carries a voice between separate first-plus-last generations, since that mode takes no audio reference, and a character's voice preset cannot reach it either. Describing the voice identically in every prompt kept these three close enough to pass as one take. If you need a guaranteed voice, the reference-to-video mode accepts referenceAudio_1, at the cost of the pinned frames.
How much does it cost? You keep your paid Google AI plan and add the flat useapi.net subscription. The whole video is 45 Flow credits, 45 cents on the Ultra $99 plan, and every image, portrait, character and the concatenate are free. See Pricing.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

Check our GitHub repo with code examples.

How to Generate Multi-Reference & Multi-Shot Video with Kling Omni via the Kling API

useapi.net — Mon, 24 Aug 2026 00:18:24 +0000

Introduction

Kling Omni is the one Kling endpoint that lets you blend several image references into a single shot and storyboard a multi-shot sequence — each scene with its own prompt and duration — in one API job, all from code against your own Kling account. Kling AI is the generative video service from Chinese short-video giant Kuaishou Technology, and useapi.net fronts it with a third-party Kling API that runs your own Kling account over a standard REST endpoint — no enterprise contract, no per-call billing from us. This guide covers Omni specifically. For plain text-to-video and start/end-frame image-to-video, see the core Kling tutorial.

What Omni does

POST /videos/omni is a single endpoint that selects a workflow from the inputs you pass. Three of those are what set Omni apart from the plain text-to-video and frames endpoints:

Multi-image reference — pass up to 7 reference images (image_1…image_7) and weave each into the prompt with @image_1, @image_2, … syntax, so one shot can combine a character, a prop, and a background.
Video Elements — reusable saved character/object references created once with POST /elements, then dropped into any later generation as @element_1 (or @object_1). Images and elements share the same pool of 7 slots, so the combined total can't exceed 7.
Multi-shot (v3 only) — split one video into 2–6 sequential shots, each with its own shot_N_prompt and shot_N_duration, for storytelling with scene cuts in a single job.

Omni also handles a frames workflow (frame_start/frame_end) and a video-reference/transform workflow (video_1), but those overlap with the core endpoints — this guide focuses on the multi-reference, Video Elements, and multi-shot paths.

Two model versions exist, picked with omni_version: v3 (the default) and o1. Multi-shot and VIDEO-type elements are v3 only — o1 supports IMAGE elements and single-clip durations of 3–10s, while v3 runs 3–15s. Pick quality with mode: std (720p), pro (1080p), or 4k (v3 only). Durations of 7s and up, counts of 2–4, and 4K need a VIP Kling plan.

Pricing

You keep your normal Kling website subscription and add a single flat $15/month to useapi.net that covers API access to every supported service, with no per-generation surcharge from us — see the core Kling tutorial's pricing section and the Kling API overview live cost calculator.

This is the consumer-account route. Kuaishou's official Kling API bills per generation at developer rates on a separate developer account, while useapi.net automates the consumer account you already pay for at the website subscription price.

Generate a multi-reference video in two API calls

You need a useapi.net API token and a connected Kling account — export the token so the curl examples below run as-is:

export USEAPI_TOKEN="user:1234-..."

Generation is asynchronous — the create call returns a task object immediately, then you poll until the video is ready.

First, upload each reference image with POST /assets (raw bytes, an image Content-Type, max 10 MB, 300px minimum). It returns a Kling-hosted url:

curl "https://api.useapi.net/v1/kling/assets/?email=user@example.com" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: image/jpeg" \
  --data-binary @character.jpg

{
  "status": 3,
  "url": "https://s21-kling.klingai.com/....jpg",
  "fileName": "abc123def456789.jpg"
}

1. Submit the job — POST https://api.useapi.net/v1/kling/videos/omni. Pass each uploaded URL as image_1, image_2, … and reference them by name in the prompt:

curl -X POST "https://api.useapi.net/v1/kling/videos/omni" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "A woman @image_1 walking through the garden @image_2, holding the lantern @image_3, camera pushing in slowly",
    "omni_version": "v3",
    "mode": "pro",
    "aspect_ratio": "16:9",
    "duration": "5",
    "image_1": "https://s21-kling.klingai.com/.../character.jpg",
    "image_2": "https://s21-kling.klingai.com/.../garden.jpg",
    "image_3": "https://s21-kling.klingai.com/.../lantern.jpg"
  }'

The response returns immediately with a task object of type m2v_omni_video. The task id you poll on is task.id (a number):

{
  "task": {
    "id": 123456789,
    "type": "m2v_omni_video",
    "status": 5,
    "status_name": "submitted",
    "status_final": false
  },
  "works": [],
  "status": 5,
  "status_name": "submitted",
  "status_final": false,
  "message": ""
}

The email field is required in the body only when you have more than one Kling account configured. A 500 from Kling almost always means a content-moderation rejection rather than a server fault — read the error text (the message field is generic and often misleading) to tell them apart. prompt maxes out at 1700 characters and references each input by its @-name (@image_1, @element_1, @video_1).

2. Poll for the result — GET https://api.useapi.net/v1/kling/tasks/{task_id}:

curl "https://api.useapi.net/v1/kling/tasks/123456789?email=user@example.com" \
  -H "Authorization: Bearer $USEAPI_TOKEN"

The task is done when status_final is true. Success is status_name: "succeed" (status: 99); the MP4 is in works[0].resource.resource:

{
  "status": 99,
  "status_name": "succeed",
  "status_final": true,
  "works": [
    {
      "workId": 123456789,
      "status_name": "succeed",
      "resource": {
        "resource": "https://s21-kling.klingai.com/....mp4",
        "height": 720,
        "width": 1280,
        "duration": 5041
      }
    }
  ]
}

The MP4 at works[0].resource.resource is watermarked. To get the clean, non-watermarked master, take the workId from the works array and call GET /assets/download — it returns a cdnUrl to the file (a single workId plus a single fileTypes value yields a direct MP4 link, otherwise a .zip):

curl "https://api.useapi.net/v1/kling/assets/download?email=user@example.com&workIds=123456789&fileTypes=MP4" \
  -H "Authorization: Bearer $USEAPI_TOKEN"

On the poll, a 404 means the task was deleted, failed at moderation, or your Kling account ran out of credits — check your balance at GET /accounts/email. Prefer not to poll? Pass a replyUrl in the create body to receive a webhook callback when the task completes.

Multi-shot sequences (v3)

Instead of one prompt and one duration, v3 Omni lets you storyboard a single video as 2–6 sequential shots, each with its own shot_N_prompt and matching shot_N_duration. The total of all durations must land between 3 and 15 seconds, shots must be sequential with no gaps (shot_1 + shot_2, never shot_1 + shot_3), and the multi-shot parameters cannot be combined with prompt or duration. Each shot prompt can carry the same @image_N / @element_N references as a single-shot job, so a recurring character holds across cuts:

curl -X POST "https://api.useapi.net/v1/kling/videos/omni" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "omni_version": "v3",
    "mode": "pro",
    "aspect_ratio": "16:9",
    "image_1": "https://s21-kling.klingai.com/.../kitchen.jpg",
    "shot_1_prompt": "Cinematic medium shot, a chef standing behind a stainless steel stove @image_1, focused intently on the pan",
    "shot_1_duration": "3",
    "shot_2_prompt": "Continuous scene at the same stove @image_1. The chef flips the food high into the air, then turns sharply to the camera",
    "shot_2_duration": "3"
  }'

The response, polling, and clean download are identical to the single-shot flow above — the create call returns the same task.id, and the finished sequence comes back as one MP4 in works[0].

Video Elements

A Video Element is a reusable character or object reference you create once and drop into any later Omni job by ID — so the same face, costume, or prop stays consistent across separate generations without re-uploading and re-describing it each time. Create one with POST /elements. There are two types:

IMAGE elements — built from a coverImage URL (optionally with extra angle views, or AI-generated multi-angle views via generateViews). Usable in both o1 and v3 Omni.
VIDEO elements — built from a video URL (mp4, minimum 3 seconds, auto-trimmed to 8s), which captures motion and can carry a voice. VIDEO elements work only in v3 Omni.

Upload the cover image (or clip) with POST /assets first, then register the element:

curl -X POST "https://api.useapi.net/v1/kling/elements" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "FashionLady",
    "coverImage": "https://s21-kling.klingai.com/.../character.jpg",
    "tag": "character"
  }'

It returns a generated element id (a 5-character random suffix is appended to your name):

{
  "elements": [
    {
      "id": "u_123456789012345",
      "name": "FashionLady ABC12",
      "description": "Elegant woman in red dress"
    }
  ],
  "count": 1
}

The tag is one of character, animal, prop, costume, scene, effect, or others (from GET /elements/tags) — leave it out to have Kling auto-detect it. List your saved elements anytime with GET /elements. Then reference the element by ID in any Omni prompt, alone or alongside images (remember the combined 7-slot limit):

curl -X POST "https://api.useapi.net/v1/kling/videos/omni" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "Character @element_1 walking through a sunlit garden, smiling",
    "omni_version": "v3",
    "mode": "std",
    "duration": "5",
    "element_1": "u_123456789012345"
  }'

A character-tagged element can also carry a voice — an official voice name/ID from GET /elements/voices, or a 5–60s mp4 to clone one from. For a VIDEO element, voice is extracted automatically from the source clip when it has 5–60 seconds of audio.

Batch script

Finding the right shot takes many attempts, and running them by hand is tedious. The Node.js script reads a list of prompts from prompts.json, submits each one to the Omni endpoint — single-shot (prompt + image/element refs) or multi-shot (shots) — then polls every task until it is final and downloads the finished MP4, preferring the clean, non-watermarked master via GET /assets/download and falling back to the watermarked works[0].resource.resource if needed. So you can queue a batch and come back to the winners.

You need Node.js v21 or newer. Put prompts.json and kling-omni.mjs in the same folder and run node ./kling-omni.mjs API_TOKEN EMAIL, where API_TOKEN is your useapi.net API token and EMAIL is your connected Kling account email. The script looks the account up by email automatically. Pass image URLs (already uploaded via POST /assets) as image_1…image_7, and saved element IDs as element_1…element_7.

Examples

The clips below are real Kling Omni generations produced through this Kling API, straight from our blog walkthroughs.

Multi-shot v3 — text-only 2-shot sequence via `POST /videos/omni` (`omni_version: v3`)

— from Kling v3: Multi-Shot Storytelling

Video Elements v3 — two VIDEO elements + a background image, 720p ~11s via `POST /videos/omni` (`omni_version: v3`, `mode: std`)

— from Kling v3: 4K Resolution and Video Elements

Frequently asked questions

What is Kling Omni? Omni is a single Kling video endpoint, POST /videos/omni, that picks a workflow from your inputs: blend up to 7 image references (@image_1…@image_7), reuse saved Video Elements (@element_1…), storyboard a v3 multi-shot sequence, run a start/end-frame transition, or guide generation from a reference video. See What Omni does above.

How do I pass multiple image references to Kling? Upload each image with POST /assets, pass the returned URLs as image_1, image_2, … (up to 7), and reference them in the prompt with @image_1, @image_2, etc. Images and saved elements share the same pool of 7 slots, so the combined total can't exceed 7. See Generate a multi-reference video in two API calls above.

What are Video Elements and how do I reuse a character? A Video Element is a reusable character/object reference created once with POST /elements — IMAGE elements from a coverImage, or VIDEO elements from an mp4 clip (v3 only). It returns an id like u_123…, which you then drop into any Omni prompt as @element_1. See Video Elements above.

How does Kling multi-shot work? On v3 Omni, set shot_1_prompt/shot_1_duration through shot_6_prompt/shot_6_duration (minimum 2 shots) instead of a single prompt/duration. Shots must be sequential with no gaps and the total duration must be 3–15 seconds. See Multi-shot sequences (v3) above.

Which Omni version supports multi-shot and Video Elements? Both are v3 only (the default). The o1 version supports IMAGE elements and single clips of 3–10s, while v3 adds VIDEO elements, multi-shot, 4k mode, and 3–15s durations. See What Omni does above.

Why is my Omni video watermarked? The MP4 at works[0].resource.resource returned by the poll is the watermarked preview. To get the clean, non-watermarked master, take the workId from the task's works array and call GET /assets/download — it returns a cdnUrl to the watermark-free file. This requires a paid Kling account. See Generate a multi-reference video in two API calls above.

My generation returns a 500 — what does that mean? Kling reuses the 500 response for content-moderation rejections as well as genuine server faults, and the generic message field rarely makes the difference clear. Read the error text instead. If a job clears creation but the poll later returns 404, the task was deleted, failed at moderation, or your account ran out of credits.

How is this different from the official Kling API? Kuaishou's official Kling API bills per generation at developer rates on a separate developer account. useapi.net instead automates your own consumer Kling account, so you generate at the website subscription price plus a flat $15/month — and the same one subscription covers 10+ other AI services.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

The full runnable example is in the kling-api GitHub repo.

Four AI Music APIs on the Same Six Briefs: TemPolor v4.8, MiniMax Music-3.0, Lyria 3.5 and Mureka V9.5

useapi.net — Thu, 20 Aug 2026 02:52:41 +0000

Introduction

TemPolor v4.8, MiniMax Music-3.0, Google's Lyria 3.5 and Mureka V9.5 — four AI music models, the same six briefs, one useapi.net API token. All 24 tracks are below, complete and unedited, each with the API call that made it. They are the same briefs we used in July, so every track here has a predecessor to compare against in AI Music APIs Compared.

TemPolor is the newcomer, and it prices differently from everything else here: a flat $30 a month for unlimited generations, where the other three charge about a cent a track. What that does to the maths is below. Listen first and judge the models yourself.

Short on time? There is also a 12-minute YouTube video — each track's loudest 30 seconds back to back, loudness-matched, with the prompt, price, and generation time on screen. The full-length tracks and the exact API call behind each one live on this page.

What changed since July

Service	July	Now	What moved
TemPolor	not in the line-up	v4.8 (songs) + i4 (instrumentals)	new to this comparison, and two new models
Mureka	V9	V9.5	new default model
Flow Music	Lyria 3 Pro	Lyria 3.5	new flagship — but 3 Pro is still the API default
MiniMax	music-2.6 via PixVerse	Music-3.0, native	new model and a new endpoint

Two of those are ordinary version bumps. The MiniMax row is not: MiniMax music is now a first-class endpoint rather than something you reach through PixVerse. Different endpoint, different billing.

TemPolor is the other exception. It is the only service here that splits the job across two endpoints with two different models: POST music/song runs v4.8 for anything with vocals, and POST music/instrumental runs i4 for anything without. Older model names still work and are routed to the new ones automatically, so existing code keeps running — but it also means you can be on the new models without noticing, or miss them by passing nothing at all.

The lineup and what it costs

Model	Through useapi.net	The vendor's own API
TemPolor v4.8 / i4	$30/month, unlimited tracks — no per-track rate at all	$0.04/call in prepaid credits — $100 to $5,000 a package, no bulk discount
MiniMax Music-3.0	$0.010/track	$0.15/composition, +$0.01 if it writes the lyrics — 15×
Google Lyria 3.5	$0.011/track	$0.08/song — 7×
Mureka V9.5	$0.0135/track	$0.03/song, and the plan starts at $1,000/month — 2×

Per-track figures are each service's highest-volume monthly plan — MiniMax Pro, Flow Music Member, Mureka Premier. Smaller plans cost more per track and still come in well under the official APIs.

TemPolor does not fit that column. Its Ultra plan sells unlimited generations for a flat monthly fee, and our submit responses came back with "unlimited": true and no remaining-track count. Its cost per track is therefore just $30 divided by however many you make — a cent at a hundred a month, a tenth of a cent at a thousand. Prices exclude VAT.

Going direct is a different shape of cost. TemPolor's own API charges $0.04 a call for the same model, paid in credits bought up front. Those credits cost $0.01 each at every package size — $100 buys 10,000, $5,000 buys 500,000 — so volume never makes a track cheaper.

What the bigger packages buy is concurrency: five requests at a time at $100, ten at $500, twenty at $5,000, and the tier does not stack if you buy the same package twice. The Ultra subscription this article ran on allows ten concurrent generations, unlimited, for $30 a month.

One naming quirk if you compare the two routes: the model this article calls v4.8 is tempolor-latest there. i4 keeps its name.

These services all sell two things: a consumer subscription, priced per month with a generous allowance, and an official API, priced per call or per credit. The API rate is several times the effective per-song rate of the subscription, because that is where the margin lives.

useapi.net connects to the account you already have. You subscribe to TemPolor, Mureka, MiniMax or Flow Music the way any user would, and we expose that account as a REST API — so generation is billed at consumer-subscription rates, not API rates. On top of that, useapi.net is a flat $15/month for every service we support, with no per-generation surcharge.

Every call in this article used a single token against all four services — no per-service SDK, no separate API keys, no four sets of billing to reconcile. The same token reaches every other service we support: video, image, speech, face swap.

How we tested

Same six briefs as July, copied from the published article rather than retyped — two instrumentals, two vocal tracks where the model writes its own lyrics, and two where we supply both the lyrics and an elaborate style brief. Every model in a round receives the identical text, so what you hear is the model. TemPolor's bodies were copied byte for byte from the MiniMax ones, which take the same prompt and lyrics fields.

Generation times are wall-clock through the API at 15-second polling granularity, so treat them as ballpark rather than benchmark. Where a model titles its own output we kept the title, spelling and stray spaces included. Tracks are complete and unedited, exactly as each API returned them: Mureka two songs per generation as 192 kbps MP3, MiniMax one at 256 kbps, TemPolor one at 48 kHz and roughly 230–320 kbps, and Flow Music an A/B pair as lossless 48 kHz WAV, transcoded to 320 kbps MP3 only so a browser can play it. Flow's lossless originals come straight from the API.

One thing to expect if you build against TemPolor: a song job reports status_final: true and COMPLETED a few seconds before its file is fetchable, and GET music/download/job_id answers 404 in the gap. Three of our four vocal rounds hit it on the first attempt. Treat that 404 as "not yet" and keep polling rather than as a failure.

Part 1 — Instrumentals

Round 1: big-band noir

Big-band noir action theme, around 60 seconds. Walking upright-bass and
baritone-sax groove, snapping rimshot snare, stabbing trumpet-section hits,
vibraphone glints, surf-twang guitar riff over a chromatic minor progression,
roaring full-brass climax with crash cymbals. Retro secret-agent swagger,
punchy analog-style production.

TemPolor i4 — "Generic Rhythm Tune" · 3:09 · generated in 103 s

curl --location 'https://api.useapi.net/v1/tempolor/music/instrumental' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model_instrumental="i4"' \
--form 'prompt="Big-band noir action theme, around 60 seconds. Walking upright-bass and baritone-sax groove, snapping rimshot snare, stabbing trumpet-section hits, vibraphone glints, surf-twang guitar riff over a chromatic minor progression, roaring full-brass climax with crash cymbals. Retro secret-agent swagger, punchy analog-style production."'

🎧 Play on useapi.net

Mureka V9.5 — "Brass Knuckle Boulevard" · 3:19 · generated in 86 s

curl --location 'https://api.useapi.net/v1/mureka/music/create-instrumental' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="V9.5"' \
--form 'prompt="Big-band noir action theme, around 60 seconds. Walking upright-bass and baritone-sax groove, snapping rimshot snare, stabbing trumpet-section hits, vibraphone glints, surf-twang guitar riff over a chromatic minor progression, roaring full-brass climax with crash cymbals. Retro secret-agent swagger, punchy analog-style production."'

🎧 Play on useapi.net

MiniMax Music-3.0 — "Midnight Operative" · 2:51 · generated in 182 s

curl --location 'https://api.useapi.net/v1/minimax/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="music-3.0"' \
--form 'prompt="Big-band noir action theme, around 60 seconds. Walking upright-bass and baritone-sax groove, snapping rimshot snare, stabbing trumpet-section hits, vibraphone glints, surf-twang guitar riff over a chromatic minor progression, roaring full-brass climax with crash cymbals. Retro secret-agent swagger, punchy analog-style production."' \
--form 'instrumental="true"'

🎧 Play on useapi.net

Lyria 3.5 — "Double Cross" · 2:52 · generated in 47 s

curl --location 'https://api.useapi.net/v1/flowmusic/music' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="lyria-3.5"' \
--form 'instrumental="true"' \
--form 'prompt="Big-band noir action theme, around 60 seconds. Walking upright-bass and baritone-sax groove, snapping rimshot snare, stabbing trumpet-section hits, vibraphone glints, surf-twang guitar riff over a chromatic minor progression, roaring full-brass climax with crash cymbals. Retro secret-agent swagger, punchy analog-style production."'

🎧 Play on useapi.net

The July versions of this brief are on the previous page if you want them back to back.

Round 2: liquid drum and bass

Liquid drum and bass instrumental, around 60 seconds. Fast rolling breakbeats
at 174 BPM, deep warm sub-bass, lush atmospheric pads, delicate piano motif,
airy wordless vocal-chop textures, smooth builds into a euphoric flowing drop.
Silky, melodic, late-night energy, pristine modern production.

TemPolor i4 — " Generic Rhythm Beat" · 3:30 · generated in 69 s

curl --location 'https://api.useapi.net/v1/tempolor/music/instrumental' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model_instrumental="i4"' \
--form 'prompt="Liquid drum and bass instrumental, around 60 seconds. Fast rolling breakbeats at 174 BPM, deep warm sub-bass, lush atmospheric pads, delicate piano motif, airy wordless vocal-chop textures, smooth builds into a euphoric flowing drop. Silky, melodic, late-night energy, pristine modern production."'

🎧 Play on useapi.net

Mureka V9.5 — "Piano Afterglow" · 3:22 · generated in 85 s

curl --location 'https://api.useapi.net/v1/mureka/music/create-instrumental' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="V9.5"' \
--form 'prompt="Liquid drum and bass instrumental, around 60 seconds. Fast rolling breakbeats at 174 BPM, deep warm sub-bass, lush atmospheric pads, delicate piano motif, airy wordless vocal-chop textures, smooth builds into a euphoric flowing drop. Silky, melodic, late-night energy, pristine modern production."' \
--form 'title="Take B - Liquid DnB"'

🎧 Play on useapi.net

MiniMax Music-3.0 — "Midnight Flow" · 3:45 · generated in 227 s

curl --location 'https://api.useapi.net/v1/minimax/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="music-3.0"' \
--form 'prompt="Liquid drum and bass instrumental, around 60 seconds. Fast rolling breakbeats at 174 BPM, deep warm sub-bass, lush atmospheric pads, delicate piano motif, airy wordless vocal-chop textures, smooth builds into a euphoric flowing drop. Silky, melodic, late-night energy, pristine modern production."' \
--form 'instrumental="true"'

🎧 Play on useapi.net

Lyria 3.5 — "Silken Tides" · 1:06 · generated in 59 s

curl --location 'https://api.useapi.net/v1/flowmusic/music' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="lyria-3.5"' \
--form 'prompt="Liquid drum and bass instrumental, around 60 seconds. Fast rolling breakbeats at 174 BPM, deep warm sub-bass, lush atmospheric pads, delicate piano motif, airy wordless vocal-chop textures, smooth builds into a euphoric flowing drop. Silky, melodic, late-night energy, pristine modern production."' \
--form 'instrumental="true"'

🎧 Play on useapi.net

None of the four APIs exposes a duration parameter — every brief asked for "around 60 seconds" in prose. Lyria was the only one that ever took the hint, and only here. Lyria 3 Pro answered this same prompt in July at 2:46.

TemPolor's instrumental endpoint also does not really name its output. It returned "Generic Rhythm Tune" and "Generic Rhythm Beat" for these two briefs — placeholders rather than titles, and not even tidy ones: the first arrives with a double space in the middle, the second with a leading space. Its song endpoint, on the next four rounds, titles properly. If you file tracks by title, generate the name yourself for instrumentals and trim whatever comes back.

Part 2 — Vocals, model writes the lyrics

Round 3: pop duet

One sentence of direction, and a deliberate capability probe — a duet only works if the model can stage two voices.

A playful pop duet between a male and a female singer about texting too late
at night, around 60 seconds, catchy back-and-forth chorus.

TemPolor v4.8 — "Text Me" · 1:51 · generated in 70 s

curl --location 'https://api.useapi.net/v1/tempolor/music/song' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model_song="v4.8"' \
--form 'prompt="A playful pop duet between a male and a female singer about texting too late at night, around 60 seconds, catchy back-and-forth chorus."'

🎧 Play on useapi.net

Listen to the first line. TemPolor wrote its own lyrics for this brief, and wrote them with the speaker tagged inline:

[Verse]
[Male] You sent "you up?" at 1:09
[Female] You said "just one more," then took your time
[Male] I saw your typing, then it slipped away
[Female] I dropped my phone and missed your "hey"

Its singer strips [Verse] and [Chorus] as structure, but an inline [Male] is not a tag it recognises — so it sings the words "male" and "female" as part of the line. The two halves of the same model disagree about its own format. MiniMax met exactly the same problem on this brief and got away with it by writing (Male) and (Female) in round brackets, which its singer does drop.

The lesson generalises past this one model. Role labels inside lyrics are not a portable API feature, so if you are supplying lyrics yourself and want two voices, check what a given model does with your bracket style before you trust it — or leave the labels out and describe the duet in the style prompt instead.

Mureka V9.5 — "Are You Still Awake?" · 3:05 · generated in 88 s

curl --location 'https://api.useapi.net/v1/mureka/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="V9.5"' \
--form 'prompt="A playful pop duet between a male and a female singer about texting too late at night, around 60 seconds, catchy back-and-forth chorus."'

🎧 Play on useapi.net

MiniMax Music-3.0 — "Three Little Dots" · 2:31 · generated in 146 s

curl --location 'https://api.useapi.net/v1/minimax/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="music-3.0"' \
--form 'prompt="A playful pop duet between a male and a female singer about texting too late at night, around 60 seconds, catchy back-and-forth chorus."'

🎧 Play on useapi.net

MiniMax named it after the typing indicator, and its (Male) / (Female) tags stayed out of the vocal.

Lyria 3.5 — "You Go First" · 1:59 · generated in 74 s

curl --location 'https://api.useapi.net/v1/flowmusic/music' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="lyria-3.5"' \
--form 'prompt="A playful pop duet between a male and a female singer about texting too late at night, around 60 seconds, catchy back-and-forth chorus."'

🎧 Play on useapi.net

Lyria wrote its own two-voice call and response, with the answering lines in parentheses.

Round 4: Spanish flamenco-pop

The language probe — the prompt is English, the song should not be:

A fiery flamenco-pop song in Spanish about summer rain, around 60 seconds,
passionate female vocals.

TemPolor v4.8 — "Lluvia" · 2:58 · generated in 102 s

curl --location 'https://api.useapi.net/v1/tempolor/music/song' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model_song="v4.8"' \
--form 'prompt="A fiery flamenco-pop song in Spanish about summer rain, around 60 seconds, passionate female vocals."'

🎧 Play on useapi.net

Mureka V9.5 — "Agua de Julio" · 3:32 · generated in 87 s

curl --location 'https://api.useapi.net/v1/mureka/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="V9.5"' \
--form 'prompt="A fiery flamenco-pop song in Spanish about summer rain, around 60 seconds, passionate female vocals."'

🎧 Play on useapi.net

MiniMax Music-3.0 — "Tacones de Lluvia" · 3:07 · generated in 171 s

curl --location 'https://api.useapi.net/v1/minimax/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="music-3.0"' \
--form 'prompt="A fiery flamenco-pop song in Spanish about summer rain, around 60 seconds, passionate female vocals."'

🎧 Play on useapi.net

Lyria 3.5 — "Lluvia de Verano" · 2:33 · generated in 73 s

curl --location 'https://api.useapi.net/v1/flowmusic/music' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="lyria-3.5"' \
--form 'prompt="A fiery flamenco-pop song in Spanish about summer rain, around 60 seconds, passionate female vocals."'

🎧 Play on useapi.net

All four wrote Spanish lyrics from an English prompt, and all four titled the track in Spanish without being asked. TemPolor's v4.8 is documented as covering 30+ languages. One Spanish brief does not test thirty of them, but it answered this one in the right language and named it "Lluvia" unprompted.

Part 3 — Vocals, our lyrics and an elaborate style brief

For the final format each model receives the same original lyrics plus a detailed production brief. Mureka takes these through music/create-advanced (lyrics + a comma-separated style desc + vocal_gender), while MiniMax, Flow Music and TemPolor all take a lyrics field next to the style prompt, so the same body works on three of the four unchanged.

Round 5: deep house

[Verse]
Two shadows in a velvet room
Moving slow to a midnight tune
Say it soft, we do not need more
The bass is talking through the floor

[Chorus]
Keep it low light, low light, hold me near
All we need is the bassline here
Round and round till the morning shows
Low light, low light, do not let go
(do not let go, do not let go)

Deep house at 121 BPM. Rolling sub-bass groove, shuffled hi-hats, warm Rhodes
chord stabs, subtle chopped vocal echoes, smoky intimate female lead vocal,
understated hypnotic late-night club production.

TemPolor v4.8 — "Low Light" · 1:28 · generated in 51 s

curl --location 'https://api.useapi.net/v1/tempolor/music/song' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model_song="v4.8"' \
--form 'prompt="Deep house at 121 BPM. Rolling sub-bass groove, shuffled hi-hats, warm Rhodes chord stabs, subtle chopped vocal echoes, smoky intimate female lead vocal, understated hypnotic late-night club production."' \
--form 'lyrics="[Verse]
Two shadows in a velvet room
Moving slow to a midnight tune
Say it soft, we do not need more
The bass is talking through the floor

[Chorus]
Keep it low light, low light, hold me near
All we need is the bassline here
Round and round till the morning shows
Low light, low light, do not let go
(do not let go, do not let go)"'

🎧 Play on useapi.net

Mureka V9.5 — "Low Light" · 3:09 · generated in 101 s

curl --location 'https://api.useapi.net/v1/mureka/music/create-advanced' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="V9.5"' \
--form 'lyrics="[Verse]
Two shadows in a velvet room
Moving slow to a midnight tune
Say it soft, we do not need more
The bass is talking through the floor

[Chorus]
Keep it low light, low light, hold me near
All we need is the bassline here
Round and round till the morning shows
Low light, low light, do not let go
(do not let go, do not let go)"' \
--form 'desc="deep house, late night, smoky, hypnotic, groovy, 121 bpm, female vocals, warm, understated club production"' \
--form 'vocal_gender="female"'

🎧 Play on useapi.net

MiniMax Music-3.0 — "Low Light" · 2:47 · generated in 135 s

curl --location 'https://api.useapi.net/v1/minimax/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="music-3.0"' \
--form 'prompt="Deep house at 121 BPM. Rolling sub-bass groove, shuffled hi-hats, warm Rhodes chord stabs, subtle chopped vocal echoes, smoky intimate female lead vocal, understated hypnotic late-night club production."' \
--form 'lyrics="[Verse]
Two shadows in a velvet room
Moving slow to a midnight tune
Say it soft, we do not need more
The bass is talking through the floor

[Chorus]
Keep it low light, low light, hold me near
All we need is the bassline here
Round and round till the morning shows
Low light, low light, do not let go
(do not let go, do not let go)"'

🎧 Play on useapi.net

Lyria 3.5 — "Velvet Room" · 2:26 · generated in 43 s

curl --location 'https://api.useapi.net/v1/flowmusic/music' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="lyria-3.5"' \
--form 'prompt="Deep house at 121 BPM. Rolling sub-bass groove, shuffled hi-hats, warm Rhodes chord stabs, subtle chopped vocal echoes, smoky intimate female lead vocal, understated hypnotic late-night club production."' \
--form 'lyrics="[Verse]
Two shadows in a velvet room
Moving slow to a midnight tune
Say it soft, we do not need more
The bass is talking through the floor

[Chorus]
Keep it low light, low light, hold me near
All we need is the bassline here
Round and round till the morning shows
Low light, low light, do not let go
(do not let go, do not let go)"'

🎧 Play on useapi.net

Round 6: jazzy house duet

The hardest brief of the set — a two-voice arrangement with live-jazz layers over a house groove, voice roles tagged in the lyrics. The tags are a hint, not an API parameter, so listen for who actually trades verses and who flattens it into one singer.

[Verse 1 - male]
Blue hour settles on the avenue
Trumpet smoke and a table for two
I talk in rhythm, you answer in rhyme
We swing together on borrowed time

[Verse 2 - female]
You bring the thunder of a double bass
I bring the shimmer of a satin dress
Piano is asking what the night is about
We are the answer when the lights go out

[Chorus - both]
Meet me in the blue hour, where the bass goes slow (blue hour, blue hour)
Sax on the corner and the lights burn low (burn low, burn low)
You take the high note, I will hold the ground
Blue hour, blue hour - do not stop the sound

TemPolor v4.8 — "Blue Hour" · 2:09 · generated in 69 s

curl --location 'https://api.useapi.net/v1/tempolor/music/song' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model_song="v4.8"' \
--form 'prompt="Jazzy deep house duet at 120 BPM. Rolling deep-house groove with swung hi-hats, walking upright-bass motif over the sub, smoky muted trumpet and tenor-sax licks answering the vocals, extended jazz piano chords, a male and a female lead trading verses and joining in harmony on the chorus, small backing ensemble echoing the hook, warm vinyl texture, sophisticated late-night club production."' \
--form 'lyrics="[Verse 1 - male]
Blue hour settles on the avenue
Trumpet smoke and a table for two
I talk in rhythm, you answer in rhyme
We swing together on borrowed time

[Verse 2 - female]
You bring the thunder of a double bass
I bring the shimmer of a satin dress
Piano is asking what the night is about
We are the answer when the lights go out

[Chorus - both]
Meet me in the blue hour, where the bass goes slow (blue hour, blue hour)
Sax on the corner and the lights burn low (burn low, burn low)
You take the high note, I will hold the ground
Blue hour, blue hour - do not stop the sound"'

🎧 Play on useapi.net

Mureka V9.5 — "Blue Hour" · 3:12 · generated in 70 s

curl --location 'https://api.useapi.net/v1/mureka/music/create-advanced' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="V9.5"' \
--form 'lyrics="[Verse 1 - male]
Blue hour settles on the avenue
Trumpet smoke and a table for two
I talk in rhythm, you answer in rhyme
We swing together on borrowed time

[Verse 2 - female]
You bring the thunder of a double bass
I bring the shimmer of a satin dress
Piano is asking what the night is about
We are the answer when the lights go out

[Chorus - both]
Meet me in the blue hour, where the bass goes slow (blue hour, blue hour)
Sax on the corner and the lights burn low (burn low, burn low)
You take the high note, I will hold the ground
Blue hour, blue hour - do not stop the sound"' \
--form 'desc="jazzy deep house, jazz house, duet, male and female vocals, group backing vocals, muted trumpet, saxophone, upright bass, swung, 120 bpm, late night, warm sophisticated production"'

🎧 Play on useapi.net

MiniMax Music-3.0 — "Blue Hour" · 2:18 · generated in 134 s

curl --location 'https://api.useapi.net/v1/minimax/music/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="music-3.0"' \
--form 'prompt="Jazzy deep house duet at 120 BPM. Rolling deep-house groove with swung hi-hats, walking upright-bass motif over the sub, smoky muted trumpet and tenor-sax licks answering the vocals, extended jazz piano chords, a male and a female lead trading verses and joining in harmony on the chorus, small backing ensemble echoing the hook, warm vinyl texture, sophisticated late-night club production."' \
--form 'lyrics="[Verse 1 - male]
Blue hour settles on the avenue
Trumpet smoke and a table for two
I talk in rhythm, you answer in rhyme
We swing together on borrowed time

[Verse 2 - female]
You bring the thunder of a double bass
I bring the shimmer of a satin dress
Piano is asking what the night is about
We are the answer when the lights go out

[Chorus - both]
Meet me in the blue hour, where the bass goes slow (blue hour, blue hour)
Sax on the corner and the lights burn low (burn low, burn low)
You take the high note, I will hold the ground
Blue hour, blue hour - do not stop the sound"'

🎧 Play on useapi.net

Lyria 3.5 — "Blue Hour" · 2:34 · generated in 58 s

curl --location 'https://api.useapi.net/v1/flowmusic/music' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="lyria-3.5"' \
--form 'prompt="Jazzy deep house duet at 120 BPM. Rolling deep-house groove with swung hi-hats, walking upright-bass motif over the sub, smoky muted trumpet and tenor-sax licks answering the vocals, extended jazz piano chords, a male and a female lead trading verses and joining in harmony on the chorus, small backing ensemble echoing the hook, warm vinyl texture, sophisticated late-night club production."' \
--form 'lyrics="[Verse 1 - male]
Blue hour settles on the avenue
Trumpet smoke and a table for two
I talk in rhythm, you answer in rhyme
We swing together on borrowed time

[Verse 2 - female]
You bring the thunder of a double bass
I bring the shimmer of a satin dress
Piano is asking what the night is about
We are the answer when the lights go out

[Chorus - both]
Meet me in the blue hour, where the bass goes slow (blue hour, blue hour)
Sax on the corner and the lights burn low (burn low, burn low)
You take the high note, I will hold the ground
Blue hour, blue hour - do not stop the sound"'

🎧 Play on useapi.net

With the lyrics supplied, all four took the title from the chorus hook. Note that this round's role labels sit inside the section headers — [Verse 1 - male] — rather than on their own at the head of a line, which is the form models treat as structure.

FAQ

What does a track cost through useapi.net? On three of the four, about a cent: MiniMax $0.010 a track, Lyria through Flow Music $0.011, and Mureka $0.0135, each on that service's highest-volume monthly plan. TemPolor has no per-track rate — its Ultra plan is a flat $30 a month for unlimited generations, so the more you generate the less each track costs.

Which one is cheapest depends on volume, then? Yes. Like for like — the same TemPolor model on its own API at $0.04 a call — the flat $30 plan breaks even at 750 tracks a month and wins above it. Because TemPolor's credits cost the same at every package size, that break-even does not move as you scale up. Against the other services here, $30 buys roughly 2,200 tracks at Mureka's rate or 3,000 at MiniMax's. Add useapi.net's flat $15/month once, whichever you pick, and nothing per generation.

Why is it cheaper than the official APIs? Because you bring your own account. useapi.net exposes the consumer subscription you already pay for as a REST API, so generation is billed at subscription rates rather than per-call API rates. useapi.net itself is a flat $15/month for every supported service, with no per-generation surcharge.

Can I run all four from one account? Yes. One useapi.net token reaches TemPolor, Mureka, MiniMax and Flow Music — every call in this article was made with a single token — plus every other service we support. You still hold your own account with each service.

How do I pick TemPolor's newest models? They are already the defaults: v4.8 on POST music/song via model_song, and i4 on POST music/instrumental via model_instrumental. Older model names are still accepted and routed to the new ones, so you cannot accidentally stay on an old model — but pass the parameter explicitly if you want your code to state which model produced a track.

Why does TemPolor have two endpoints when the others have one? Because vocals and instrumentals are different models there — v4.8 and i4. Pick the endpoint by whether you want singing, not by a flag. The other three take an instrumental parameter on a single endpoint instead.

How do I pick Lyria 3.5 rather than 3 Pro? Pass model: "lyria-3.5" on POST /music. lyria-3-pro remains the default, so omitting the parameter renders the older model.

Do any of these APIs let me set the track length? No. None of the four exposes a duration parameter. The amount of lyrics you supply is the only lever — more words generally means a longer song.

Can I get more than one track per call? On MiniMax, yes — quantity up to 3, billed as three songs. Mureka returns two songs per generation and Flow Music an A/B pair, both included in the one generation. TemPolor returned one track per call throughout this test, and allows up to ten generations running at once.

Can I put role labels like [Male] in the lyrics I send? You can send them, but no API here treats them as a parameter, and what happens next differs per model — TemPolor sings inline [Male] out loud, MiniMax drops (Male). Describe the duet in the style prompt and test your bracket style before relying on it. See round 3.

MiniMax H3 vs Seedance 2.0: 2K with native audio, head to head on the same omni references

useapi.net — Fri, 31 Jul 2026 05:04:22 +0000

MiniMax's newest video model, H3, is live on POST videos/create as model: "Hailuo-3.0". Two things separate it from everything else in the MiniMax lineup: every generation ships with native stereo audio — no separate voice or music pass — and it renders 2K at 24 fps by default, with the short edge at 1440 pixels (2560×1440 at 16:9, 1440×2560 at 9:16).

H3 is a general-purpose multimodal model rather than a set of task-specific modes. The same endpoint takes a prompt alone, a start frame, or a tray of image, video, and audio references, and the request shape is the one Seedance 2.0 already uses — set model, resolution, and duration directly.

The three clips below are the actual output of a single 4-second run at each mode, unedited. Turn the sound on — the audio in all three is generated by the model.

The last of the three is a direct comparison: it reuses the exact image, video, and audio references and the exact prompt from our Seedance 2.0 omni post, so H3 and the three Seedance 2.0 tiers can be judged on identical input rather than on separate demos.

Text to video

No references at all, just a prompt. aspectRatio is explicit here, though H3 defaults to 16:9 when no reference is supplied.

Aerial push over jagged sea cliffs at dawn. Waves crash white against black rock far below as mist burns off in the first gold light. A lone seabird crosses frame. Cinematic wide-angle, natural color grade. Crashing surf, wind, and distant gull calls.

Your browser does not support the video element.

The surf, wind, and gulls are all model-generated — nothing was added to the audio track.

A start frame, with speech

Supply one image as the first frame and let the prompt drive what happens next. The portrait below is the same reference used in our Seedance 2.0 omni post, so the two models can be compared on identical input.

fileID — the start frame. Upload it with POST /files first.

She looks into the camera, smiles, and says warmly: "You made it — I was starting to think you'd stood me up." Warm bar lights bokeh behind her, subtle handheld drift on the camera. Low jazz and crowd murmur sit under her voice.

Your browser does not support the video element.

Spell the spoken line out in the prompt. H3 renders the delivery from the text, so quoting the line and saying it is spoken is what puts a voice on the track.

Omni references — image, video, and audio together

The same three references from the Seedance 2.0 omni post, and the same prompt, run through H3 for a direct comparison. H3 accepts up to 9 images, 3 videos, and 3 audio clips, capped at 12 files in total, each addressed by an @-tag in the prompt.

Your browser does not support the video element.

@video1 — the scene, rope, and motion to keep. Its face is blurred on purpose: a clear real face in a video reference is rejected by the content filter, and the identity comes from @image1 anyway.

🎧 Play on useapi.net

@audio1 — the speech to drive, about 2 seconds. Together with the portrait above as @image1, that is the full tray.

Replace the woman in @video1 with @image1, keeping the same outdoor scene, rope, and jumping movements. She speaks out loud, clearly saying "Count me jumping" in the voice from @audio1, audible throughout the clip.

Your browser does not support the video element.

No aspectRatio was sent here. With references attached H3 defaults to Auto and matched the 9:16 shape of the inputs on its own.

What it costs

H3 bills 12 credits per second of output at 2K. A video reference is charged on its own length, at the same rate, on top of the generated clip — image and audio references add nothing per second.

Request	Credits
4-second clip, no references	48
4-second clip, image reference	48
4-second clip, 10-second video reference	168
15-second clip, no references	180

That surcharge is worth planning around: trim a video reference to the shortest clip that carries the motion you need, since its full length is billed whatever the output duration.

Where H3 fits

Seedance 2.0 still reaches 4K and runs cheaper per second at 480p and 720p, so it remains the better pick for volume iteration and for the highest resolution. H3 is the one to choose when you want 2K with sound in a single call — the native audio track removes an entire step from the pipeline, and 24 fps at 1440 is a materially different starting point for finishing work.

See the POST videos/create reference for every parameter, the reference-media limits, and a live Try-It console.

How to Generate AI Music with the Mureka API (Suno / Udio Alternative)

useapi.net — Mon, 27 Jul 2026 00:39:22 +0000

Introduction

Mureka AI turns a single text prompt into a complete, produced song — a direct Suno and Udio alternative — and the useapi.net Mureka API drives your own Mureka account over plain REST, with its latest V9 default and flagship O2 models, at your normal consumer subscription price. Mureka is the music generator from Chinese tech company Kunlun Tech, built on its proprietary Mureka models (mureka-9, mureka-o2). One generation returns two complete songs — vocals, melody, and arrangement — and from there you can write your own lyrics, force an instrumental, seed a track from a reference song or melody, extend a song, or clone a voice for text-to-speech. useapi.net runs all of it from one API token.

Pricing

You keep your normal Mureka web subscription for the underlying credits, plus a single flat $15/month to useapi.net that covers API access to every supported service — there is no per-song surcharge from us. Mureka's own subscription plans:

Plan	Price	Songs / month	Cost per song
Pro	$9/month	up to 500	~1.8¢
Premier	$27/month	up to 2,000	~1.35¢

Song and speech allowances are shared (songs or speech minutes), lyrics generation is free, and downloading instrumentals and stems is free. The newer V9 and O2 models consume more Gold per song, so real counts run lower — roughly 250–400 songs on V9/O2. Pro/Premier members can buy additional Gold packs (price shown at checkout).

This is the consumer-account route. Mureka also runs an official developer API, but it works differently — it is sold as prepaid top-ups with 12-month validity, consumed per call rather than billed monthly, and the ladder runs from $30 (1 concurrent request) up to $30,000 (150 concurrent), with the $1,000 tier giving 5 concurrent. Generation there runs ~$0.045/song on V8/V9/O2 (~$0.03 on the older V7.6), plus extra charges for lyrics generation and stem export. useapi.net instead automates the consumer plan you already pay for, so a $27 Premier plan plus the flat $15 covers far more generation than the official API's entry tier. For other music options under the same subscription, see the Flow Music API tutorial (Google's Lyria 3 Pro) and the TemPolor API.

Supported models

Pick a model per request with the model field on POST /music/create. Omit it and the API applies its default, V9.

Model	Engine	What it's for
`V9` (default)	`mureka-9`	Latest generation, used when `model` is omitted — advanced prompt control and studio-grade sound.
`O2`	`mureka-o2`	Flagship high-end model — best audio quality and musical coherence, with strong multilingual vocals.
`V8`	`mureka-8.0.1`	Prior-generation general-purpose model.
`V7.6`	`mureka-7.6`	Earlier model kept for compatibility — it was the default before V9.

Older aliases (V7.5, V7, O1, V6) are still accepted but redirect to current engines, so prefer the four models above. Every model returns finished MP3 audio — there is no MIDI or sheet-music output. The same account and token also drive Mureka's text-to-speech and voice cloning, covered below.

Generate a song in two API calls

You need a useapi.net API token and a connected Mureka account — export the token so the curl examples below run as-is:

export USEAPI_TOKEN="user:1234-..."

The most reliable workflow is asynchronous: the create call returns a jobid immediately, then you poll until both songs are ready.

1. Submit the job — POST https://api.useapi.net/v1/mureka/music/create:

curl -X POST "https://api.useapi.net/v1/mureka/music/create" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "account": "12345678901234",
    "prompt": "upbeat synthwave with bright female vocals about chasing a sunrise, 110 bpm",
    "model": "V9",
    "async": true
  }'

The account is your Mureka account id, and it is optional when you have a single account configured (the API auto-selects it). In async mode the response returns immediately with 201 Created, a jobid, and status: "created":

{
  "jobid": "j0622184042017475905m-u1234-a12345678901234-bot:mureka",
  "verb": "music/create",
  "jobType": "music",
  "status": "created",
  "created": "2026-06-22T18:40:42.412Z",
  "request": {
    "account": "12345678901234",
    "prompt": "upbeat synthwave with bright female vocals about chasing a sunrise, 110 bpm",
    "model": "V9",
    "async": true
  }
}

2. Poll for the result — GET https://api.useapi.net/v1/mureka/jobs/{jobid} (the jobid ends in bot:mureka, so URL-encode it in the path — it contains a :):

curl "https://api.useapi.net/v1/mureka/jobs/JOBID" \
  -H "Authorization: Bearer $USEAPI_TOKEN"

When status is completed, the finished audio lives in response.songs[]. Each generation returns two songs (an A/B pair, version "1" and "2"), and each carries an mp3_url plus cover art:

{
  "jobid": "j0622184042017475905m-u1234-a12345678901234-bot:mureka",
  "status": "completed",
  "response": {
    "feed_id": 11223344,
    "state": 3,
    "songs": [
      {
        "song_id": "user:1234-mureka:123456789-song:33445566",
        "title": "Chasing the Sunrise",
        "version": "1",
        "duration_milliseconds": 173040,
        "mp3_url": "https://static-cos.mureka.ai/cos-prod/...1.mp3",
        "cover": "https://static-cos.mureka.ai/cos-prod/...1.png"
      },
      {
        "song_id": "user:1234-mureka:123456789-song:33445567",
        "title": "Chasing the Sunrise",
        "version": "2",
        "duration_milliseconds": 178910,
        "mp3_url": "https://static-cos.mureka.ai/cos-prod/...2.mp3",
        "cover": "https://static-cos.mureka.ai/cos-prod/...2.png"
      }
    ]
  }
}

A song typically finishes in 30–90 seconds (about 45 on average), and an account supports up to 10 parallel generations. Songs can run up to about 5 minutes depending on the lyrics. Jobs are retained for 7 days, so you can re-fetch a finished result any time in that window. Prefer not to poll? Pass a replyUrl in the create body to receive a webhook callback when the job completes (the payload matches the poll response). A job can come back status: "failed" (e.g. a content-policy rejection), and the create call may return 402 (useapi.net subscription expired) / 412 (out of Mureka credits), 429 (rate limit, or all 10 generation slots busy), or 596 (account session error) — see the endpoint docs for the full list.

Write your own lyrics, or force an instrumental

The plain prompt on POST /music/create lets the model write the lyrics for you. To supply your own words, switch to POST /music/create-advanced — lyrics is required there (use [Verse] / [Chorus] tags, 50–5,000 chars), and desc steers the style with a comma-separated list of genres, moods, and vocal direction. vocal_gender forces male or female:

# Your own lyrics → forces a vocal render of those exact words
curl -X POST "https://api.useapi.net/v1/mureka/music/create-advanced" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "lyrics": "[Verse 1]\nCity lights blur in the rain\n[Chorus]\nAnd I will wait for the morning sun",
    "desc": "soulful R&B, lush keys, 90 bpm",
    "vocal_gender": "female",
    "model": "V9",
    "async": true
  }'

For a purely instrumental track, use POST /music/create-instrumental — just a prompt describing the music, no lyrics:

curl -X POST "https://api.useapi.net/v1/mureka/music/create-instrumental" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "prompt": "ethereal post-rock, cascading delay guitars, orchestral swells building to a cinematic crescendo", "model": "O2", "async": true }'

All three endpoints return the same jobid you poll on, and any resulting song_id can be lengthened with POST /music/extend (pass the song_id plus new lyrics) or reworked with POST /music/regenerate.

Reference tracks, melody seeding, and custom vocals

POST /music/create-advanced can seed a song from your own audio instead of a text description, using three reference inputs (each mutually exclusive with desc):

Parameter	Seed from	Set it up with
`ref_id`	A reference track whose style inspires the song	POST /files (upload), GET /music/refs (browse)
`motif_id`	A melody idea that kickstarts the song	POST /files/motif
`vocal_id`	A specific singing voice, including your own uploaded vocal	POST /files/vocal, GET /music/vocals

This is how you make Mureka sing your custom lyrics in a voice you supply — upload an MP3 vocal sample, take the returned id, and pass it as vocal_id alongside your lyrics. The same Pro-tier reference features (custom soundtracks and melodies) that the Mureka website offers are all reachable from the API.

Clone a voice and generate multi-speaker speech

The same account also drives Mureka's text-to-speech. POST /speech turns text into audio with either a preset or a cloned voice, and unlike music generation it has no concurrency (generation-slot) limit. Clone a voice first from an MP3 sample with POST /speech/voice, then pass the returned id as the voice_id parameter with your text:

curl -X POST "https://api.useapi.net/v1/mureka/speech" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{ "text": "Hello, this is a cloned voice speaking.", "voice_id": 12345, "async": true }'

For a back-and-forth dialogue, drop text/voice_id and send a conversation array instead — each turn names its own voice_id, so two or more cloned voices can talk to each other in one render:

curl -X POST "https://api.useapi.net/v1/mureka/speech" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "conversation": "[{\"voice_id\": 12345, \"text\": \"Is there anything I can help you with?\"}, {\"voice_id\": 67890, \"text\": \"Yes — read me back the weather.\"}]",
    "async": true
  }'

Speech jobs poll the same way, with the finished MP3 in response.mp3_url. See GET /speech/voices for the available voices, including the ones you have cloned.

Batch-generate with a script

Finding the right track takes many attempts, and running them by hand is tedious. The Node.js script reads a list of prompts from prompts.json, submits each one to POST /music/create in async mode, polls every job until it is done, then downloads both MP3s from each generation — so you can queue a batch and come back to a folder of finished songs.

You need Node.js v21 or newer. Put prompts.json and mureka.mjs in the same folder and run node ./mureka.mjs API_TOKEN, where API_TOKEN is your useapi.net API token. With more than one Mureka account configured, pass the account id (or its Google email) as a second argument — with a single account it is auto-selected. Each entry's prompt is required, and the optional model maps straight to the create endpoint.

Examples

These are real songs and speech generated through the Mureka API, straight from our blog walkthroughs — each from the request shapes shown above. Full prompts and curl for each are in the linked posts.

Astro Gardener — `V9`, female-vocal indie pop with custom lyrics (`create-advanced`)

🎧 Play on useapi.net

— from Mureka: V9 Now the Default Music Model

Cascading Sunrise — `V9` instrumental (`create-instrumental`)

🎧 Play on useapi.net

— from Mureka: V9 Now the Default Music Model

Galactic Odyssey — flagship `O2` instrumental (`create-instrumental`)

🎧 Play on useapi.net

— from Mureka: O2 and V7.6 Music Models

Pirate Programmer — `V7.6`, male-vocal sea-shanty-meets-electronic (`create-advanced`)

🎧 Play on useapi.net

— from Mureka: O2 and V7.6 Music Models

Custom vocal clone — an uploaded voice singing custom lyrics over a reference track (`create-advanced` with `vocal_id` + `ref_id`)

— from Mureka: Custom Vocals

Multi-speaker speech — two cloned voices in one conversation (`POST /speech`)

🎧 Play on useapi.net

— from Mureka: Speech Generation and Voice Cloning

Frequently asked questions

Is the Mureka API a Suno or Udio alternative? Yes. Like Suno and Udio, Mureka generates full songs with vocals from a single text prompt — and the useapi.net Mureka API drives your own Mureka account over plain REST, with the latest V9 default and flagship O2 models, all under one flat $15/month useapi.net subscription. See Supported models above.

Is there an official Mureka API? Yes. Mureka runs an official developer API, sold as prepaid top-ups (12-month validity) consumed per call — from $30 (1 concurrent request) up to $30,000 (150), with the $1,000 tier giving 5 concurrent. Generation runs ~$0.045/song on V8/V9/O2 (~$0.03 on V7.6), plus charges for lyrics and stems. useapi.net is the consumer-account route, automating the $9–$27/month Mureka web plan you already pay for. See Pricing above.

Which models can I use? V9 (default), the flagship O2, V8, and V7.6. Omit the model field and the API uses V9. Older aliases (V7.5, V7, O1, V6) still work but redirect to current engines. See Supported models above.

Can I write my own lyrics or generate an instrumental? Yes. Use POST /music/create-advanced with [Verse]/[Chorus]-tagged lyrics to force a vocal render of your exact words, or POST /music/create-instrumental for a lyrics-free track. See Write your own lyrics, or force an instrumental above.

Can I clone a voice or generate speech? Yes. POST /speech/voice clones a voice from an MP3 sample, then POST /speech renders text — including multi-speaker conversation dialogues — with no concurrency (generation-slot) limit. See Clone a voice and generate multi-speaker speech above.

What do I get back, and in what format? Every music generation returns two complete songs (an A/B pair), each as an mp3_url with cover art, up to about 5 minutes long. The output is finished MP3 audio — there is no MIDI or sheet-music export. See Generate a song in two API calls above.

How much does it cost? You keep your normal Mureka web plan for credits (~1.35–1.8¢ per song), plus a flat $15/month to useapi.net for API access to all services, with no per-song surcharge. See Pricing above.

How is this different from the official Mureka API? The official Mureka API is sold as prepaid top-ups (12-month validity) consumed per call, from $30 up to $30,000, charging ~$0.045/song on V8/V9/O2 (~$0.03 on V7.6) plus lyric and stem fees. useapi.net instead automates your own consumer Mureka account, so you generate at the Mureka website subscription price plus a flat $15/month — and the same one subscription covers other AI services too, including the Flow Music and TemPolor music APIs.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

The full runnable example is in the mureka-api GitHub repo.

Seedance 2.0 API Pricing: The Cheapest Ways to Run It, Compared

useapi.net — Tue, 21 Jul 2026 20:08:55 +0000

Introduction

Seedance 2.0 is the video model everyone is asking for right now, and the cheapest verified way to run its 1080p API is through useapi.net at about $1.20 a clip — below ByteDance's own official rate of $1.87 and below every third-party reseller we could verify. We pulled live pricing on July 21, 2026 from ByteDance's official channels, sixteen third-party vendors, and our own four routes, then normalized everything to one honest unit: the dollar cost of a 5-second, 1080p, text-to-video Seedance 2.0 clip.

It is worth being upfront about where we sit before the numbers. useapi.net is a bring-your-own-account layer, not an official ByteDance reseller. We drive your own PixVerse, Runway, Dreamina, or MiniMax account over a single REST API, so you spend those platforms' consumer credits instead of the metered developer-API rate. That is exactly why the numbers come out lower — and we will show the math rather than ask you to trust it. Prices move fast in this market, so every figure below carries its date, and you should re-check the source before committing spend.

What a Seedance 2.0 clip costs at the source

ByteDance sells Seedance 2.0 through two official channels — BytePlus ModelArk internationally and Volcano Engine in China. Both bill per million tokens, where a clip's token count is width × height × 24fps × seconds ÷ 1024. A 1080p second is 48,600 tokens, so a 5-second clip is 243,000 tokens at 1080p's published rate of $7.7 per million — $1.87. The per-million rate itself changes with resolution — lower at 720p, and lower still at 4K — which is why a 4K clip is not simply four times a 1080p one. Normalized, per official BytePlus rates (July 2026):

Resolution	5-second clip	10-second clip
480p	$0.35	$0.70
720p	$0.76	$1.51
1080p	$1.87	$3.74
4K	$3.89	$7.78

Seedance 2.0 Fast and Mini are cheaper but cap at 720p — there is no official 1080p or 4K on those tiers. Audio generation is included at no extra cost on the 2.0 family. Volcano runs roughly 8–13% cheaper than BytePlus at current exchange rates, but it requires Chinese real-name and face verification, so it is effectively China-only.

The sticker price is not the whole cost. Activating Seedance 2.0 on BytePlus requires a prepaid resource pack — minimum $30.10, non-refundable, expiring after 90 days — and there is no free trial for the 2.0 models. Individual accounts are throttled to 3 concurrent tasks (180 requests/minute), and 4K is capped at a single concurrent task for everyone. None of that makes the official API a bad deal. It sets the bar every alternative should be measured against.

Four ways to run Seedance 2.0 through useapi.net

One useapi.net token drives Seedance 2.0 on four different platforms, each with a different strength. All four sit under the same flat $15/month that covers every service we support, on top of the underlying account plan.

Route	Seedance models	5s 1080p	Stands out for
PixVerse	2.0, Fast, Mini	$1.20	cheapest 1080p, 4K, and real human faces
Dreamina	2.0, Fast, Mini, 1.5 Pro, 1.0 family	~$1.57 (Canada accounts)	widest model range, the only route with Seedance 1.x
MiniMax	2.0, Fast, Mini	~$1.90	4K plus a rich reference-image workflow
Runway	seedance-2, seedance-2-fast	metered	real faces with lighter moderation, plus a legacy unlimited mode

The PixVerse route is the price leader. Its cost calculator prices a 1080p Seedance 2.0 second at 60 credits, so a 5-second clip is 300 credits — $1.20 on the $60/month Premium plan at $0.004 per credit. That is the generation cost per clip. Counter-intuitively, the Premium plan beats the pricier Ultra plan for single clips, because Ultra's 40%-off third-party discount is outweighed by its worse per-credit rate. PixVerse also reaches 2160p (4K) on Seedance 2.0, at about $2.40 for a 5-second clip — below both MiniMax's 4K and the official rate. It handles real human faces too, one of the reference workflows the official API restricts (more on that below).

The Runway route earns its place on real faces, not price. It generates Seedance 2.0 at up to 1080p on metered credits, with notably lighter moderation of real people than the other routes. It also carries a legacy unlimited mode — on older Runway Unlimited plan accounts that already have Explore mode enabled, exploreMode runs Seedance at 480p and 720p without spending credits. Runway no longer enables Explore mode on new accounts, though, and where it does run it is slow — one clip at a time, roughly 10 to 15 minutes each. Treat it as a perk if you already have it, not a reason to pick Runway.

The trade-offs are real and worth naming. Our routes automate consumer accounts, so concurrency is account-bound (typically 5–10 parallel jobs, versus the official API's per-key limits), Dreamina's 1080p and 4K are locked to Canada-region accounts, and Runway's unlimited Explore mode is a legacy perk that new accounts can no longer enable. What you get in return is a materially lower per-clip cost and one token across all four platforms.

The real-faces gap

Price is not the only place the official API is restrictive. Seedance 2.0 rejects reference images and videos that contain real human faces, unless the face comes from the model's own recent output, a preset digital character, or a verified real-person asset. Doing real-person reference work at any scale means ByteDance's enterprise "Advanced Creation Rights" — $14,000 a year for a million assets, $42,000 for five million, corporate verification required.

Three of our routes soften this. The PixVerse route accepts real human faces, the Runway route documents real-face support with notably lighter content moderation, and the MiniMax route accepts real faces in image references (video references with recognizable faces are still moderated). For advertising, UGC, and creator workflows that revolve around a real person, that difference matters more than a few cents per clip — and it does not require the official API's enterprise contract.

The wider market, priced honestly

Most third-party Seedance vendors are reselling ByteDance's own API with a margin on top, and the honest ones cluster at 20–60% above official. A few reference points, priced live in July 2026 for a 5-second 1080p clip:

Vendor	5s 1080p	Notes
useapi.net (PixVerse route)	$1.20	consumer-account credits
Segmind	$1.70	the only reseller we verified below official
Official BytePlus	$1.87	the yardstick
Replicate	$2.25	"Official" model badge, transparent per-second pricing
fal.ai	$3.41	the market's default, and the most expensive standard 2.0 we found

That last row is worth sitting with. fal.ai is where a lot of developers reach for Seedance first, and at $14 per million tokens it charges more than 1.8× ByteDance's own $7.7 — an ~82% premium for the same model.

You will also see sticker prices that look far lower than official. Every one we chased down decomposed into something other than a cheaper Seedance 2.0. A few patterns to watch for when a price looks too good:

Flat per-second rates that hide a cross-subsidy — a single "$0.11/second at any resolution" number is cheap at 1080p but quietly more expensive than official at 480p, and it usually leans on venture subsidy that will not last.
"Turbo" or "accelerated" variants that are not the standard model — one vendor advertises a 1080p clip at $0.75 under a model name ByteDance does not publish, while its actual standard Seedance 2.0 costs four times that on the same site.
Prices quoted at 720p when you wanted 1080p — several "cheapest Seedance" pages have no 1080p tier at all, so the low number is a 720p rate, and a homepage claim of "30–70% cheaper" turns into roughly 9% more than official once you match resolutions.
"Unlimited" plans capped below 1080p — an unlimited tier that only reaches 720p, runs on a slow queue, and expires in a fixed window is a different product than an unlimited 1080p API.
Credits with no published dollar value — if a page prices generations in credits but never tells you what a credit costs, you cannot compare it, and that is usually the point.

None of this means those vendors are dishonest — it means resolution, model variant, and fine print have to match before a comparison means anything. When they do, the verified market has no one meaningfully below official except Segmind at a thin margin, and our PixVerse route beneath that.

Show me the API call

Every price above is one REST call away. Generate a 1080p Seedance 2.0 clip through the PixVerse route:

curl --location 'https://api.useapi.net/v2/pixverse/videos/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="seedance-2.0"' \
--form 'prompt="A lone lighthouse on a storm-battered cliff at dusk, waves exploding against the rocks, beam sweeping through the rain, cinematic wide shot."' \
--form 'quality="1080p"' \
--form 'duration="5"' \
--form 'aspect_ratio="16:9"'

Or, if your Runway account already has Explore mode enabled — Runway no longer offers it on new accounts — the same shape runs unlimited at 720p by adding exploreMode, one slow clip at a time:

curl --location 'https://api.useapi.net/v1/runwayml/videos/create' \
--header 'Authorization: Bearer user:12345-…' \
--form 'model="seedance-2"' \
--form 'text_prompt="A lone lighthouse on a storm-battered cliff at dusk, waves exploding against the rocks, beam sweeping through the rain, cinematic wide shot."' \
--form 'resolution="720p"' \
--form 'duration="5"' \
--form 'exploreMode="true"'

Both return a job id to poll (or a replyUrl webhook when the clip is ready). The full parameter set — reference images, first and last frames, aspect ratios, and audio — lives on each endpoint page linked in the table above.

Which route should you use for Seedance 2.0?

Cheapest 1080p, or you need 2160p — the PixVerse route, at $1.20 for a 5-second 1080p clip.
Real human faces — the PixVerse, MiniMax, or Runway route, which avoid the official API's enterprise-gated face restriction.
Widest model choice, or you specifically want Seedance 1.5 Pro / 1.0 — the Dreamina route, the only one carrying the 1.x family.
4K with a deep reference-image workflow — the MiniMax route, or PixVerse for 4K at the lower price.

Examples

Real Seedance 2.0 output from each route, reused from our blog posts — the same model, run on different platforms. Three of these lean on real human faces, the workflow the official API gates behind an enterprise contract.

PixVerse · Seedance 2.0 (Standard), 1080p — a real face from a photo dropped into reference footage and given speech from a reference audio clip, all in one omni request. From PixVerse Seedance 2.0: Mini vs Fast vs Standard.

Your browser does not support the video element.

MiniMax · Seedance 2.0 (Standard), 1080p — the same real-face omni workflow on the MiniMax route. From MiniMax Seedance 2.0: Standard vs Fast vs Mini.

Your browser does not support the video element.

Runway · Seedance 2.0, 15 seconds, 9:16 — a real face animated on the Runway route, which applies lighter moderation on real people. From Seedance 2.0: Using Real Faces.

Your browser does not support the video element.

Dreamina · Seedance 2.0, text-to-video with lip-sync from a prompt alone. From Dreamina: Seedance 2.0 Video Generation.

Your browser does not support the video element.

FAQ

Is useapi.net an official Seedance API? No. useapi.net is a REST layer that automates your own PixVerse, Runway, Dreamina, or MiniMax account, each of which runs Seedance 2.0. You spend those platforms' consumer credits rather than ByteDance's metered developer-API rate, which is why the per-clip cost comes out lower.

What is the cheapest way to run Seedance 2.0 at 1080p? Of everything we verified in July 2026, the PixVerse route at about $1.20 per 5-second clip, below ByteDance's official $1.87 and below every third-party reseller. See the pricing table.

Can I generate Seedance 2.0 videos with real human faces? The official API restricts real-face reference inputs to verified assets or an enterprise contract starting at $14,000 a year. The PixVerse and MiniMax routes accept real faces (MiniMax in image references), and the Runway route documents real-face support with notably lighter moderation.

How does useapi.net charge less than ByteDance's own API? The official developer API meters by token for scale customers. Consumer plans on PixVerse, Runway, Dreamina, and MiniMax price the same generation lower because they are sold to creators. useapi.net automates one of those consumer accounts over REST, so you get the consumer rate through an API, plus a flat $15/month across every service.

Are the sub-dollar Seedance prices I see elsewhere real? Usually not for a standard 1080p Seedance 2.0 clip. They tend to be 720p rates, subsidized flat-rate plans, distilled "turbo" variants, or credits with no published dollar value — see the wider market section.

Ready to try it? Pick a route — PixVerse for the cheapest 1080p and real faces, MiniMax for 4K with a reference workflow — connect the account, and a useapi.net subscription drives all four with one token. Prices in this article were current on July 21, 2026, and this market moves quickly, so confirm the live rate before you scale.

AI Music APIs Compared: Mureka vs MiniMax vs ElevenLabs vs Lyria 3 Pro

useapi.net — Fri, 17 Jul 2026 03:10:32 +0000

Introduction

We gave the same briefs to five AI music models — Mureka O2, Mureka V9, MiniMax music-2.6, ElevenLabs music-v1, and Google's Lyria 3 Pro — through useapi.net, and published every result: 30 full tracks below, with real prices, generation times, and the quirks we hit along the way. Three test formats, two rounds each: pure instrumentals, vocals where the model writes its own lyrics, and vocals singing lyrics we wrote. Every model in a round gets the identical brief, so what you hear is the model, not the prompt.

The five models arrive through three services — Mureka directly (the strongest Suno and Udio alternative in the group), MiniMax and ElevenLabs through PixVerse, and Lyria 3 Pro through Flow Music. One useapi.net token drives all three. If you want the copy-paste quick starts instead of a listening test, see the Mureka tutorial and the Flow Music tutorial.

Short on time? The whole test is also a 15-minute YouTube video — a skimmable cut with every track's best 30 seconds back to back, loudness-matched, with the prompt, model, cost, and generation time on screen. The full-length tracks and the exact API call behind each one live on this page.

The lineup and what a track costs

Model	useapi.net API	Cost per track	Official API
Mureka O2	Mureka API	$0.0135	Official Mureka API starts at $1,000/month at 3¢/song
Mureka V9	Mureka API	$0.0135	same as above
MiniMax music-2.6	PixVerse API	$0.16	~$0.15/track up to 5 min
ElevenLabs music-v1	PixVerse API	$0.60	$0.15/min, about $0.75 for 5 min
Lyria 3 Pro	Flow Music API	~$0.008–0.013	$0.08/song on Google's official API

Per-track figures are what the underlying account plan works out to — Mureka Premier at $27/month for 2,000 songs, PixVerse credits at $0.004 each on the $60/month Premium plan, Flow Music paid plans. On top of the account you pay useapi.net's flat $15/month for API access to every supported service, with no per-generation surcharge. The standouts are at the extremes: Lyria through Flow Music runs six to ten times cheaper than Google's own API for the same model, and Mureka through useapi.net is about half the official per-song rate with no $1,000/month floor. For ElevenLabs the PixVerse route is about running everything through one account rather than price — at the one-minute tracks it likes to produce, the official API is cheaper.

One playback note before you listen: Flow Music offers a lossless WAV alongside its m4a, so the Lyria players below are WAVs — the rest are the mp3s their APIs return.

How we tested

Every prompt asked for "around 60 seconds" (spoiler: only one model listened — see what we learned). Generation times are wall-clock through the API, polled at 15-second intervals, so treat them as ballpark. Where a model titles its own output we kept the title. Tracks are complete and unedited.

Part 1 — Instrumentals

Round 1: big-band noir

A retro secret-agent groove — walking bass, stabbing brass, surf-twang guitar. Rhythm feel and section writing are hard to fake:

Big-band noir action theme, around 60 seconds. Walking upright-bass and
baritone-sax groove, snapping rimshot snare, stabbing trumpet-section hits,
vibraphone glints, surf-twang guitar riff over a chromatic minor progression,
roaring full-brass climax with crash cymbals. Retro secret-agent swagger,
punchy analog-style production.

Mureka O2 — “The Dead Drop” · 3:19 · generated in 119 s

🎧 Play on useapi.net

Mureka V9 — “Dead End Alley” · 3:28 · generated in 85 s

🎧 Play on useapi.net

MiniMax music-2.6 · 2:11 · generated in 99 s

🎧 Play on useapi.net

ElevenLabs music-v1 · 1:00 · generated in 16 s

🎧 Play on useapi.net

Lyria 3 Pro — “The Shadow Trail” · 2:59 · generated in 34 s

🎧 Play on useapi.net

Note the lengths: same brief, same "around 60 seconds" request — ElevenLabs delivered exactly 1:00, everyone else played on.

Round 2: liquid drum and bass

Maximum contrast with round 1 — a steady 174 BPM groove exposes rhythm programming and texture instead of dynamics:

Liquid drum and bass instrumental, around 60 seconds. Fast rolling breakbeats
at 174 BPM, deep warm sub-bass, lush atmospheric pads, delicate piano motif,
airy wordless vocal-chop textures, smooth builds into a euphoric flowing drop.
Silky, melodic, late-night energy, pristine modern production.

Mureka O2 — “Silver Lining” · 3:19 · generated in 86 s

🎧 Play on useapi.net

Mureka V9 — “Silken Motif” · 2:31 · generated in 67 s

🎧 Play on useapi.net

MiniMax music-2.6 · 2:22 · generated in 119 s

🎧 Play on useapi.net

ElevenLabs music-v1 · 1:00 · generated in 16 s

🎧 Play on useapi.net

Lyria 3 Pro — “Silky Liquid Flow” · 2:54 · generated in 50 s

🎧 Play on useapi.net

Part 2 — Vocals, model writes the lyrics

Round 3: pop duet

One sentence of direction, and a deliberate capability probe — a duet only works if the model can stage two voices. That is the thing to listen for in this round, and it is the clearest differentiator of the six.

A playful pop duet between a male and a female singer about texting too late
at night, around 60 seconds, catchy back-and-forth chorus.

Mureka O2 — “Don’t Hit Send” · 3:18 · generated in 80 s

🎧 Play on useapi.net

Mureka V9 — “Midnight Dance” · 2:53 · generated in 62 s

🎧 Play on useapi.net

MiniMax music-2.6 · 2:40 · generated in 97 s

🎧 Play on useapi.net

ElevenLabs music-v1 · 1:00 · generated in 16 s

🎧 Play on useapi.net

Lyria 3 Pro — “Don’t Say Goodnight” · 2:58 · generated in 79 s

🎧 Play on useapi.net

Round 4: Spanish flamenco-pop

The language probe — the prompt is English, the song should not be:

A fiery flamenco-pop song in Spanish about summer rain, around 60 seconds,
passionate female vocals.

Mureka O2 — “El llanto del sol” · 3:59 · generated in 79 s

🎧 Play on useapi.net

Mureka V9 — “Lluvia de Verano” · 2:32 · generated in 63 s

🎧 Play on useapi.net

MiniMax music-2.6 · 2:57 · generated in 113 s

🎧 Play on useapi.net

ElevenLabs music-v1 · 1:00 · generated in 17 s

🎧 Play on useapi.net

Lyria 3 Pro — “Tierra Mojada” · 2:34 · generated in 64 s

🎧 Play on useapi.net

All five accepted the brief and every model sang in Spanish. Lyria even wrote a full Spanish lyric sheet ("Cae el peso del cielo gris / Sobre el polvo de este jardín…") and titled the track accordingly, as did both Murekas.

Part 3 — Vocals, our lyrics and an elaborate style brief

For the final format each model received the same original lyrics plus a detailed production brief. Mureka takes these through music/create-advanced (lyrics + a comma-separated style desc + vocal_gender), PixVerse and Flow Music take a lyrics field next to the style prompt.

Round 5: deep house

[Verse]
Two shadows in a velvet room
Moving slow to a midnight tune
Say it soft, we do not need more
The bass is talking through the floor

[Chorus]
Keep it low light, low light, hold me near
All we need is the bassline here
Round and round till the morning shows
Low light, low light, do not let go
(do not let go, do not let go)

Deep house at 121 BPM. Rolling sub-bass groove, shuffled hi-hats, warm Rhodes
chord stabs, subtle chopped vocal echoes, smoky intimate female lead vocal,
understated hypnotic late-night club production.

Mureka O2 — “Low Light” · 3:13 · generated in 60 s

🎧 Play on useapi.net

Mureka V9 — “Low Light” · 2:39 · generated in 57 s

🎧 Play on useapi.net

MiniMax music-2.6 · 1:06 · generated in 59 s

🎧 Play on useapi.net

ElevenLabs music-v1 · 0:40 · generated in 29 s

🎧 Play on useapi.net

Lyria 3 Pro — “Low Light (Take 2)” · 2:32 · generated in 44 s

🎧 Play on useapi.net

Round 6: jazzy house duet

[Verse 1 - male]
Blue hour settles on the avenue
Trumpet smoke and a table for two
I talk in rhythm, you answer in rhyme
We swing together on borrowed time

[Verse 2 - female]
You bring the thunder of a double bass
I bring the shimmer of a satin dress
Piano is asking what the night is about
We are the answer when the lights go out

[Chorus - both]
Meet me in the blue hour, where the bass goes slow (blue hour, blue hour)
Sax on the corner and the lights burn low (burn low, burn low)
You take the high note, I will hold the ground
Blue hour, blue hour - do not stop the sound

Jazzy deep house duet at 120 BPM. Rolling deep-house groove with swung hi-hats,
walking upright-bass motif over the sub, smoky muted trumpet and tenor-sax licks
answering the vocals, extended jazz piano chords, a male and a female lead
trading verses and joining in harmony on the chorus, small backing ensemble
echoing the hook, warm vinyl texture, sophisticated late-night club production.

Mureka O2 — “Blue Hour” · 2:49 · generated in 59 s

🎧 Play on useapi.net

Mureka V9 — “Blue Hour” · 2:27 · generated in 72 s

🎧 Play on useapi.net

MiniMax music-2.6 · 1:35 · generated in 62 s

🎧 Play on useapi.net

ElevenLabs music-v1 · 1:12 · generated in 32 s

🎧 Play on useapi.net

Lyria 3 Pro — “Blue Hour Duet (Retry)” · 3:03 · generated in 60 s

🎧 Play on useapi.net

What we learned

Nobody has a duration parameter, and only one model listens to the prompt

None of the five APIs exposes a track-length parameter. Every brief above asked for "around 60 seconds" — here is what that request was worth across all our runs, including test takes not published here:

Model	Asked for ~60 s, got	Verdict
ElevenLabs music-v1	1:00 on every prompt-only run	the only model that treats it as a spec
Lyria 3 Pro	~0:59 on 2 prompts, 2:30–3:03 on the rest	occasionally
MiniMax music-2.6	1:20–2:57	ignores it
Mureka O2 / V9	2:30–4:00	ignores it

The lever that actually works is lyric length. With our short one-verse-one-chorus lyrics in part 3, ElevenLabs came in at 0:40, MiniMax dropped to 1:06, and Mureka fell from its usual 3–4 minutes to as short as 0:50 in test takes. If you need a fixed-length cue, write fixed-length lyrics — or trim in post.

ElevenLabs refuses named artists and works, the rest do not

In side experiments we prompted "in the spirit of the James Bond GoldenEye title sequence" and, separately, "in the spirit of Pendulum and The Prodigy". ElevenLabs music-v1 failed both generations at the service level — evidently a copyrighted-reference filter that catches both named works and named artists. Mureka O2, Mureka V9, MiniMax music-2.6, and Lyria 3 Pro accepted and generated both times. Style-descriptive prompts (like round 1's unnamed secret-agent brief) sail through everywhere.

Speed, titles, and what lands in your bucket

ElevenLabs is in a different speed class — 16 to 32 seconds per track in our runs, wall-clock through the API. Lyria followed at 34–79 s, with Mureka V9 at 57–85 s and both Mureka O2 and MiniMax spanning 59–119 s. Mureka and Lyria name their own tracks — every title above is model-written, and in part 3 both pulled the name from our lyric hook, quirky suffixes included ("Take 2" and "Retry" on the Lyria tracks are Lyria's own). The PixVerse-routed models return untitled files. Mureka gives you two variants per generation at 320 kbps, PixVerse one track at 192 kbps, and Flow Music usually an A/B pair with a lossless WAV available — the only lossless source in the group.

FAQ

Which AI music API is the cheapest? Per track in this lineup: Lyria 3 Pro via Flow Music at roughly $0.008–0.013, then Mureka at $0.0135 (Premier plan rate, two variants per generation), MiniMax music-2.6 at $0.16, and ElevenLabs music-v1 at $0.60 via PixVerse. See the lineup for how these compare with the official APIs.

Can I control how long the generated song is? Not with a parameter — no API here has one. Asking in the prompt only works reliably on ElevenLabs. The dependable lever is lyric length: short lyrics produce short songs on every model that accepts custom lyrics.

Can I bring my own lyrics? Yes on all five models: Mureka through music/create-advanced, MiniMax and ElevenLabs through the PixVerse lyrics field, and Lyria through the Flow Music lyrics field. Tag structure with [Verse] and [Chorus].

Will these models imitate a named artist or song? ElevenLabs refuses generation when the prompt names a work or artist — we hit its filter twice. The other four accepted named-reference prompts in our tests. Describing a style without names works on all five.

Ready to run your own listening test? The whole experiment is one script — github.com/useapi/ai-music-comparison sends any brief (the six above are included) to all five models and downloads every track. Then dig deeper with the Mureka and Flow Music tutorials, the PixVerse music endpoint, and a useapi.net subscription — one token, all five models.

Nano Banana 2 Lite vs 2 vs Pro: Google Flow Image Models Compared

useapi.net — Thu, 02 Jul 2026 22:22:01 +0000

Introduction

Google Flow ships three Nano Banana image models — nano-banana-2-lite, nano-banana-2, and nano-banana-pro — and you can drive all three from one prompt through the useapi.net Google Flow API. You use your own Google Flow account, so there's no Google Cloud project, no API key, and no per-image metering — image generation is included on any Google AI plan, even a free one, where the official Gemini API bills $0.034–$0.134 per image.

This guide runs the exact same prompt through Lite, 2, and Pro so you can see how they differ, then hands you a runnable Node.js script that generates the whole set in one command. Want a from-scratch how-to first? See the sibling image-generation tutorial.

What are the three Nano Banana models?

Pick a model per request with the model field. If you omit it, Google Flow uses nano-banana-2-lite — its default since July 2026. All three run on any Google AI subscription, including a free account.

Model id	Marketing name	Best for	References (I2I)	Speed
`nano-banana-2-lite` (default)	Nano Banana 2 Lite / Gemini 3.1 Flash-Lite Image	fast, high-volume generation on any plan	max 10	fastest
`nano-banana-2`	Nano Banana 2 / Gemini 3.1 Flash Image	balanced quality and speed	max 10	fast
`nano-banana-pro`	Nano Banana Pro / Gemini 3 Pro Image	highest fidelity and instruction-following	max 10	slower

All three accept the same aspect ratios (16:9, 4:3, 1:1, 3:4, 9:16, plus auto in image-to-image mode), count (1–4), seed for reproducibility, up to 10 reference images, inline @-mentions, and 2K/4K upscaling. The deprecated aliases imagen-4 (Google removed Imagen from Flow in July 2026) and nano-banana are still accepted and map to nano-banana-2-lite and nano-banana-2 respectively. Full per-model detail is on the POST /images reference.

Upscaling to 2K/4K

Any generated image can be upscaled with POST /images/upscale: pass its mediaGenerationId and a resolution of 2k (default) or 4k. The response is base64 in encodedImage. 2k works on any plan; 4k requires a paid Google account.

Pricing

Third-party Google Flow API by useapi.net vs. the official Gemini API — drive your own Google Flow subscription instead of metered, per-call API billing:

Model	Official Gemini API	useapi.net (Flow Pro)	useapi.net (Flow Ultra)
Veo 3.1 Fast — 8s clip	$0.80	~$0.40	~$0.10
Veo 3.1 Quality — 8s clip	$3.20	~$2.00	~$1.00
Veo 3.1 Lite — 8s clip	$0.40	~$0.20	~$0.05
Veo 3.1 Lite, lower priority	—	—	$0 (Ultra $199)
Gemini Omni Flash — 8s clip	— (Flow only)	~$0.50	~$0.25
Nano Banana Pro — per image	$0.134	included	included
Nano Banana 2 — per image	$0.067	included	included
Nano Banana 2 Lite — per image	$0.034	included	included

Free workload (no Flow credits spent)	Avg generations / day*
Veo 3.1 Lite — lower priority video	~1,000
Images — Nano Banana 2 Lite, Nano Banana 2, Nano Banana Pro	up to ~500

*Averages from the busiest real accounts — not guarantees. They are set by Google's own allowances and fluctuate with demand. Applies only to the free lower-priority video queue and image generation. Credit-metered models (Veo 3.1 Fast / Quality and Gemini Omni Flash) are not shown here — their volume is bounded by your plan's monthly Flow credits (see the credit table), not a free allowance.

Full plan and credit details are on the Google Flow overview. See the setup guide to connect your account.

The comparison prompt

A model comparison is only useful if the prompt is hard. This one is a deliberately absurd, single-frame scene that stress-tests several things at once: a named photographic style (David LaChapelle), a multi-subject spatial layout (a swarm forming an arrow in the top-left, a forced-perspective bee in the bottom-right), prop fidelity (dripping honeycomb), a specific comedic facial expression (o_O cartoon eyes), and a clashing tone — all in one cohesive image.

A visually breathtaking, beautifully absurd, single-frame photograph in the hyper-saturated,
glossy, surreal-comedy editorial style of David LaChapelle, shot at a sun-dappled backyard
apiary at golden hour on an 85mm f/1.4 lens with creamy shallow depth of field. In the center
midground, a gorgeous, radiant young woman beekeeper with an elegant figure, flawless glowing
sun-kissed skin and a warm playful smile, incongruously wearing a tiny mini bikini and a
beekeeper's veil hat pushed back off her face, stands among the wooden hive boxes on the grass and lifts a single
open beehive frame heavy with golden honeycomb, thick honey oozing and stringing down in glossy
ribbons and catching the light. She looks directly into the camera, delighted and completely
unbothered by the bees swirling around her. In the top-left of the frame, a dense swarm of bees
bunches together and organizes into a bold, unmistakable arrow / pointer shape aimed straight at
her, as if frantically calling everyone's attention. In the extreme bottom-right foreground of
the exact same scene, photobombing the lens, is a huge forced-perspective close-up of a single
frazzled worker bee, staring directly into the camera with bulging, mind-blown "o_O" cartoon
googly eyes, completely frozen in comical disbelief. Single cohesive image, seamless depth of
field, no cut-ins, warm dynamic backlighting, glistening honey highlights, crazy situational
comedy, hyper-detailed, masterpiece.

Generate across all three models

You need a useapi.net API token and a connected Google Flow account. POST /images is synchronous — one call blocks until the images are ready (typically 20–60 seconds, and the Pro model can take up to ~2 minutes under load) and returns a media[] array with one entry per image, so there's no jobid to poll.

A single call — swap model to compare:

curl -X POST "https://api.useapi.net/v1/google-flow/images" \
  -H "Authorization: Bearer $USEAPI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-pro",
    "aspectRatio": "9:16",
    "count": 4,
    "prompt": "A beekeeper at golden hour lifting a dripping honeycomb frame, a swarm forming an arrow in the sky, a forced-perspective bee with o_O googly eyes in the foreground"
  }' > result.json

# each image is at .media[N].image.generatedImage.fifeUrl
curl -o pro_1.jpg "$(jq -r '.media[0].image.generatedImage.fifeUrl' result.json)"

To run the full comparison in one command, the Node.js script reads prompts.json — one entry per model, all with the same prompt — submits each, and downloads every returned image with a model-labeled filename. You need Node.js v21+. Put both files in a folder and run node ./google-flow-nano-banana-compare.mjs API_TOKEN EMAIL, where API_TOKEN is your useapi.net API token and EMAIL is your connected Google Flow account. The script looks the account up by email and checks its health before submitting, and sends no captchaToken — the useapi.net worker solves the reCAPTCHA automatically.

Prefer to clone and run it? The complete script, a Python port, and prompts.json are on GitHub in useapi/google-flow-api.

Examples

Here is the same beekeeper prompt rendered by each model at 9:16, generated through this Google Flow API. The differences are consistent: Pro is the most polished and the most literal about the instructions, 2 is the most dynamic, and Lite is the fastest while staying close.

nano-banana-pro — the most polished and editorial, and the only model that reliably renders the veil netting, the bees crawling on the comb, and the literal o_O cartoon eyes.

nano-banana-2 — the most dynamic and energetic interpretation, with the clearest arrow-shaped swarm and lots of situational detail (work boots, bees on the frame).

nano-banana-2-lite — the fastest of the three (roughly half Pro's time in our runs) while keeping the composition, the honey drip, and the googly-eyed foreground bee.

Which Nano Banana model should you use?

Reach for nano-banana-pro when fidelity matters most — hero shots, complex multi-subject scenes, precise text or instruction-following. It is the slowest of the three.
Reach for nano-banana-2 for a strong balance of quality and speed, or when you want a more dynamic, stylized read of a prompt.
Reach for nano-banana-2-lite (the default) for iteration and high-volume batches where speed matters and the quality is still very close.

Because they share the same request shape, switching is a one-word change to model — generate a batch across all three and pick the winner, exactly like the script above does.

Run it yourself

The complete Node.js script, a Python port, and prompts.json live in useapi/google-flow-api/nano-banana-compare. Clone it, edit prompts.json, and run one command to generate the full comparison set from your own account.

Frequently asked questions

Is there a Nano Banana 2 or Nano Banana Pro API? Yes — two ways. Google offers these models on the official Gemini API, which meters per image on a Google Cloud project. Or use useapi.net's Google Flow API, which exposes the same models as nano-banana-2, nano-banana-2-lite, and nano-banana-pro on the POST /images endpoint — driving your own Google Flow account with no Cloud project and no per-image metering, on any Google AI plan including the free tier.
What is the difference between Nano Banana 2 Lite, 2, and Pro? They are three tiers of Google's Nano Banana image family. nano-banana-pro (Gemini 3 Pro Image) is the highest-fidelity model, nano-banana-2 (Gemini 3.1 Flash Image) balances quality and speed, and nano-banana-2-lite (Gemini 3.1 Flash-Lite Image) is the fastest and the default. All three accept up to 10 reference images, the auto aspect ratio in image-to-image mode, and 2K/4K upscaling.
Which Nano Banana model is best? For maximum quality, nano-banana-pro. For everyday and high-volume work, nano-banana-2-lite is the default and the fastest, with quality very close to the larger models. See Which model should you use? above.
Is it free? Image generation works on any Google AI plan, including a free account, plus a flat monthly subscription to useapi.net for API access to every service. The official Gemini API meters per image instead — see Pricing.
How much does the Nano Banana API cost? Through useapi.net there is no per-image charge — you keep your own Google AI plan (a free account works for images) and pay a flat $15/month for API access to every service. On the official Gemini API the same models meter per image: roughly $0.034 (2 Lite), $0.067 (2), and $0.134 (Pro) per 1K image — see Pricing.
Do generations get moderated? Google applies content moderation, and a rejected request returns an empty result. Moderation decisions vary between requests and between models, so retrying, editing the prompt, or switching models often succeeds — each Nano Banana model moderates a little differently.
Do I need a captcha or a Google Cloud project? No Google Cloud project. Generation requires reCAPTCHA, but you don't solve it — the useapi.net worker does it automatically, and your first Google Flow account comes with 300 free captcha credits (a one-time grant powered by CapSolver). The script here sends no captchaToken and relies on that automatic solving.

Conclusion

Visit our Discord Server or Telegram Channel for any support questions and concerns.

Check our GitHub repo with code examples.