If your product talks in a cloned voice, that clone is the hardest thing to move off a text-to-speech vendor. It lives in their account, they meter every character it says, and each new lesson, video or support call makes leaving a bit more expensive.
ElevenLabs released Eleven v4 on September 28, 2026, at $0.08 per 1,000 characters over the API. Its web app charges for the same characters in credits, at roughly double that. To compare vendors on equal terms we picked one job: a US course platform that narrates 1 million characters a month in its founder's cloned voice. We priced it on ElevenLabs and eight alternatives. It comes to $80 on Eleven v4, $49 on Cartesia, $15 on Fish Audio and $10 on Murf, though Murf only clones voices on Enterprise. This is the short version. The full comparison on DevToolLab has the complete pricing script and a screenshot of every vendor's page.
Two prices for one character at ElevenLabs
As of October 15, 2026, the ElevenLabs API pricing page lists Eleven v4 at $0.08 per 1,000 characters and Eleven v4 Turbo at $0.04. The changelog says Turbo has a median inference latency of about 100 ms. The monthly plan fee is usage you pay for up front at that same rate: $6 for Starter covers 75,000 v4 characters, $22 for Creator covers 275,000, and $99 for Pro covers 1,238,000.
Studio, the web app, is billed in credits. Creator comes with 121,000 of them, which is about 121 minutes of audio. Each extra minute costs $0.17 to $0.36 depending on your plan. ElevenLabs' own Pay As You Go docs admit the gap: "API usage has lower rates than UI usage."
Some details to watch:
- The free plan is for non-commercial use and requires attribution. A commercial license and Instant Voice Cloning start at Starter, and Professional Voice Cloning starts at Creator.
- Pay As You Go balances expire 12 months after you buy them. When your balance reaches $0 the service stops, unless Auto Top Up is on.
- Eleven v4 dropped SSML support and the Style and Speed sliders. If your pipeline relies on markup, you'll be rewriting part of it whether you stay or leave.
Pricing the same month everywhere
ElevenLabs treats roughly 1,000 characters as one minute of speech. That makes our test job about 1,000 minutes, or 16.7 hours of commercially licensed narration a month. This script prices it on the vendors with simple per-character billing:
// tts-month.mjs - one month of cloned-voice narration, priced per service.
// usage: node tts-month.mjs [characters] (rates checked October 15, 2026)
const chars = Number(process.argv[2] ?? 1_000_000)
const perMillion = (rate, free = 0) => (Math.max(0, chars - free) / 1e6) * rate
// ElevenLabs API: the $6 Starter fee is prepaid usage, Pay As You Go covers the rest.
const elevenApi = (perK, included) => 6 + (Math.max(0, chars - included) / 1000) * perK
const bill = {
"ElevenLabs v4 (API)": elevenApi(0.08, 75_000),
"ElevenLabs v4 Turbo (API)": elevenApi(0.04, 150_000),
"Cartesia Sonic-3.6 (Startup)": 49 + perMillion(45, 1_250_000), // cheapest Cartesia plan at 1M
"Fish Audio S2.1-Pro": perMillion(15),
"Murf Falcon 2": perMillion(10),
"Google Chirp 3 HD": perMillion(30, 1_000_000),
}
for (const [name, usd] of Object.entries(bill).sort((a, b) => a[1] - b[1]))
console.log(name.padEnd(30) + ("$" + usd.toFixed(2)).padStart(9))
Google Chirp 3 HD $0.00
Murf Falcon 2 $10.00
Fish Audio S2.1-Pro $15.00
ElevenLabs v4 Turbo (API) $40.00
Cartesia Sonic-3.6 (Startup) $49.00
ElevenLabs v4 (API) $80.00
That is a trimmed version of the original script. The version on DevToolLab also covers Studio credits, Deepgram, Rime and Google's custom voices. It picks the cheapest plan for whatever volume you pass in and prices voice agent minutes too. Here is the full result:
| Service | 1M characters | Cloned voice |
|---|---|---|
| Google Chirp 3 HD | $0 | Stock voices only, first 1M free each month |
| Murf Falcon 2 | $10 | Enterprise plan only |
| Fish Audio S2.1-Pro | $15 | Self-serve over the API |
| Deepgram Aura-2 | $30 | No cloning |
| Rime Mist v3 | $30 | Ordered as an enterprise project |
| ElevenLabs v4 Turbo (API) | $40 | Self-serve |
| Cartesia Sonic-3.6 (Startup) | $49 | Self-serve from the $5 Pro plan |
| Rime Coda | $50 | Ordered as an enterprise project |
| Google Instant Custom Voice | $60 | Allow-listed accounts only |
| ElevenLabs v4 (API) | $80 | Self-serve |
| ElevenLabs Studio (Pro) | $167 | Self-serve, web UI credits |
| Chatterbox or Kokoro, self-hosted | $0 plus compute | Chatterbox only |
For a support line handling 10,000 voice agent minutes a month, the agent platforms come to $800 on ElevenLabs Agents (the LLM is billed separately at cost), $750 on Deepgram's Voice Agent Standard (a standard-tier LLM is included) and $600 on Cartesia Managed Agents (plus $0.014 a minute if the call runs on a Cartesia phone number).
Check the clone column before you look at price. Google, Murf, Deepgram and Rime either use stock voices or only clone through a sales process. Of the vendors that let you clone a voice yourself, Fish Audio at $15 is the cheapest. If you want to stay on ElevenLabs, the easy saving is moving narration from Studio to the API, which takes this month from $167 down to $80.
Cartesia: one credit per character
Cartesia charges one credit per character for Sonic-3.6, a model that covers 44 languages. You can pin it to the dated snapshot sonic-3.6-2026-08-27, which will not change underneath you. Cartesia's product page says first audio arrives in under 90 ms.
Its cheapest paid plan already includes cloning. Pro costs $5 a month and gives you a commercial license plus Instant Voice Cloning from a clip of 10 to 60 seconds. Startup adds two Professional Voice Clone slots, each trained on 30 or more minutes of audio. Managed Agents run $0.06 a minute.
Costs climb once you use up your plan's credits: $65 per extra million on Pro and $45 on Startup. Concurrency is also low, with 3 parallel requests on Pro and 5 on Startup.
Pricing: Pro $5/month (100,000 credits) · Startup $49 (1.25M) · Scale $299 (8M) · proprietary
Fish Audio: priced per byte, clones at no charge
Fish Audio charges $15 per million UTF-8 bytes for S2.1-Pro, which supports 83 languages. There is no subscription or minimum spend. English text takes one byte per character, so our test month costs $15.
You create a clone with a single POST /model request that uploads your sample, and the pricing page doesn't charge for it. Until November 30, 2026, the same model is also available as s2.1-pro-free, which costs $0 under fair-use limits.
The free variant comes with no time-to-first-audio guarantee and no DPA. Accounts are limited to 5 concurrent requests until they've paid $100. Drama 3, which launched October 13, 2026, isn't on the pricing page yet. The hosted API is proprietary. The open S2 Pro weights are licensed for research and non-commercial use only, and commercial use needs a license from Fish Audio.
Murf Falcon 2: the lowest hosted rate, with a catch
Murf charges 1 cent per 1,000 characters for Falcon 2, the cheapest hosted rate we found. It publishes a live 30-day median time to first audio of 95 ms. Our test month costs $10. That covers 150+ voices in 35 languages, with endpoints in us-east, us-west and other regions. Concurrency starts at 5 in US-East and 2 everywhere else.
The catch is cloning. Murf's docs say Instant Voice Cloning "is available on the Enterprise plan only," and the cloning endpoints return 403 until your account has it. Murf also publishes a migration guide for former Play.ai users. On October 15, 2026, the play.ht domain no longer resolved.
Deepgram: an agent bill that includes the LLM
Deepgram charges $0.030 per 1,000 characters for Aura-2 and $0.045 for Flux TTS, a newer model built for voice agents. When a caller interrupts, Flux TTS reports exactly how much of the response they heard. Deepgram says output can start in as little as 80 ms.
The main reason to choose Deepgram is that the whole agent comes on one bill. On Pay As You Go, its Voice Agent API costs $0.075 a minute. That covers speech-to-text, a Standard-tier LLM such as gpt-5.4-mini, and speech output. New accounts get $200 in credit, and Enterprise customers can self-host on NVIDIA GPUs. Deepgram doesn't offer cloning, though. Its docs have no voice cloning page, and Flux TTS voices are English-only.
Rime: low latency and on-prem deployment
Rime has two models: Coda (May 2026) at $0.05 per 1,000 characters and Mist v3 (March 2026) at $0.03. The Starter plan has no monthly fee. Rime lists a P50 time to first audio of 37 ms for Mist v3 and 96 ms for Coda. On Enterprise, both can run on-premises or in your VPC through Docker Compose or Kubernetes, and Rime will sign a BAA for HIPAA workloads.
There is no instant cloning. Each clone is an enterprise project that needs 30 to 60 minutes of clean recordings. Starter allows 20 concurrent generations.
Google Cloud: a free million, then gated clones
Google Cloud Text-to-Speech doesn't charge for the first 1 million Chirp 3 HD characters each month and bills $30 per million after that. If stock voices work for you, our test month costs nothing. Cloning is less accessible. Instant Custom Voice costs $60 per million characters with no free tier. It is only open to allow-listed accounts, and the speaker must record a consent statement.
Chatterbox: free cloning, with a watermark
Chatterbox is Resemble AI's family of open models. Multilingual V3 has 0.5B parameters and covers 23+ languages. Turbo (350M) and Nano (110M) are English-only. All three clone a voice from a short reference clip.
Know the downside before you choose it: every file it produces carries Resemble's PerTh watermark, which survives MP3 compression. In return you get commercial cloning under the MIT license with no per-character fee. The README says Nano runs three times faster than real time on 8 CPU cores, so batch narration doesn't need a GPU. The repo had 26,955 stars on October 15, 2026.
Kokoro: small and permissive, but no cloning
Kokoro is an 82M-parameter model with Apache 2.0 weights, 54 voices and 8 languages. It was downloaded 10,468,205 times on Hugging Face in the month to October 15, 2026. You install it with pip install kokoro, and it runs on modest hardware. It can't clone a voice, though, so it can't read in yours. Its latest package, 0.9.4, was released April 5, 2025.
Side by side
| Service | 1M chars/month | Self-serve clone | Agent minute | Self-host |
|---|---|---|---|---|
| ElevenLabs | $80 (v4), $40 (v4 Turbo) | From $6 Starter | $0.08 + LLM | Enterprise, v2/v2.5 only |
| Cartesia | $49 | From $5 Pro | $0.06 | Enterprise |
| Fish Audio | $15 | Yes | Beta, unpriced | Research weights only |
| Murf | $10 | Enterprise | Sales only | Enterprise |
| Deepgram | $30 (Aura-2) | No | $0.075, LLM included | Enterprise |
| Rime | $30 to $50 | Enterprise | Not sold | Enterprise |
| Google Cloud | $0 stock, $60 cloned | Allow-listed | $0.06 to $0.12 | Not listed |
| Chatterbox | Hardware only | Yes | Not sold | Yes, MIT |
| Kokoro | Hardware only | No | Not sold | Yes, Apache 2.0 |
List prices as of October 15, 2026.
Picking one without moving twice
- Use your real volume. Export last month's character count from ElevenLabs usage analytics and plug it into the script. Plan thresholds change the ranking at higher volumes. If you only have a sample script, the Character Counter gives you the billable count with spaces included, and the Reading Time Calculator estimates how many minutes it runs.
- Keep the original recordings. A clone can't be exported from one vendor to another. Cartesia clones from 10 to 60 seconds of audio and Murf from up to 30 seconds, but Rime needs 30 to 60 minutes.
- Test your hardest material. Render one script full of numbers, one full of names and one long passage on your two finalists.
- Check the license on the weights, not just the code. F5-TTS code is MIT, but its weights are CC BY-NC 4.0. Fish Audio's open S2 Pro weights are also non-commercial.
-
Measure latency from your own region, for most US stacks that means
us-east-1. Every latency figure in this post comes from the vendor.
Recommendations
- Your own cloned voice for narration, on a budget: Fish Audio, $15 for the test month, with the clone created over the API.
- A real-time agent that speaks in a cloned voice: Cartesia, with cloning on the $5 plan and agents at $0.06 a minute.
- An agent with the LLM on the same bill: Deepgram's Voice Agent API at $0.075 a minute, if a stock voice is acceptable.
- HIPAA or on-prem requirements: Rime Enterprise if you want a vendor that signs a BAA, or self-hosted Chatterbox if no audio can leave your network.
- Staying on ElevenLabs: move narration from Studio to the API ($167 down to $80), then try v4 Turbo at $40.
Wrapping up
ElevenLabs is betting on v4's quality. Meanwhile Murf, Fish Audio and Rime charge $10 to $50 per million characters, against $80 on v4. OpenAI is also retiring tts-1, tts-1-hd and both gpt-4o-mini-tts snapshots on January 6, 2027, and pointing users to its Realtime API, so more teams will be comparing vendors soon. Before you renew any TTS contract, make sure you still own the source recordings you'd need to re-clone your voice somewhere else.
References
- ElevenLabs Alternatives in 2026, Priced (original article on DevToolLab)
- ElevenLabs API pricing
- Cartesia pricing
- Fish Audio pricing and rate limits
- Murf API
- Deepgram pricing
- Rime pricing
- Google Cloud Text-to-Speech pricing
- Chatterbox on GitHub
- Kokoro-82M on Hugging Face
- Best Text-to-Speech APIs (DevToolLab)
Top comments (0)