Start on Gemini 4 Hydrogen (gemini-4-hydrogen) and move up only when it fails your own tests. Hydrogen costs $1.50 per 1M input tokens and $7.50 per 1M output tokens, and Google's pricing page lists no end date for that rate.[3] Gemini 4 Carbon and Argon cost $2 and $10 through November 30, 2026, then $4 and $20 from December 1.[1] After that date a Carbon or Argon token costs 2.67 times a Hydrogen token. All three models share a 1,048,576-token window and a 65,536-token default output.[4] Below: IDs and limits, a worked monthly bill per model, curl and Python calls, access gates, a skeptical look at the MKA benchmark, and a decision table. Argon's full tier table is in our Argon pricing post.
This post reports published figures only and includes no benchmark runs of its own. Every number comes from Google's launch posts (Carbon on November 10, Hydrogen on November 12, 2026), the Gemini API and Vertex AI docs, or Artificial Analysis.[1][2]
The three models side by side
| Model | Model ID | Input / output per 1M tokens | Cache read per 1M | Default thinking level | Free tier |
|---|
| Gemini 4 Carbon | gemini-4-carbon | $2 / $10 through Nov 30, then $4 / $20 | $0.10, then $0.20 | medium | No |
| Gemini 4 Argon | gemini-4-argon | $2 / $10 through Nov 30, then $4 / $20 | $0.10, then $0.20 | medium | No |
| Gemini 4 Hydrogen | gemini-4-hydrogen | $1.50 / $7.50, flat | $0.075 | high | Yes, rate-limited |
Prices are the Standard tier on the Gemini API pricing page.[3] The output rate includes thinking tokens on all three, and Google bills every token the model processes, including tool instructions and function declarations when grounding is on.[3]
The limits are identical. The 1,048,576-token context window is shared across input, thinking and output. Maximum output is 65,536 tokens for a normal synchronous request, or up to 1,048,576 minus your input with long decode. Inputs can be text, image, video, audio or PDF, and output is text only.[4][5] All three support function calling, structured outputs, code execution, computer use and Search grounding. None supports explicit caching; implicit caching covers prompts of at least 4,096 tokens and discounts cached input by 95%.[6]
What each model is for
Carbon: the deep-verification model
Google calls Carbon its strongest model and says it is built for tasks that test the limits of a model's intelligence, such as complex multi-step coding and debugging and open-ended scientific research.[1] The cost is tokens. Carbon uses more thinking tokens than Argon to explore thoroughly and run extended verification loops, which raises per-turn token use and latency.[1][6] Artificial Analysis shows the same pattern: Carbon at High scores 60 on its Intelligence Index v5.0.1 (4th of 81 models), but needs 32.1 minutes per Index task against a 13.2-minute median and generated 205M output tokens against a 78.6M median.[14]
Argon: the efficient frontier model
Google says Argon remains the ideal choice for the majority of software engineering and knowledge work.[1] It shares Carbon's price and window but spends fewer thinking tokens per turn, so interactive coding and day-to-day feature work run faster and cheaper per turn.[6]
Hydrogen: the everyday and agent model
Hydrogen targets everyday coding, tool-using agents and parallel workers. Google's comparison table lists its speed as Fast, against Moderate for Argon and Deliberate for Carbon.[8] It adds a fourth thinking level, xhigh, and defaults to high. Google reports that Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default level against 50 for Argon, reaching 54 at xhigh.[2] Those are vendor-reported figures.
What it costs: now versus after December 1
The per-token ratios decide most of the arithmetic. Through November 30, Argon and Carbon cost 1.33 times Hydrogen on every line. From December 1 they cost 2.67 times Hydrogen. So if Hydrogen needs no more than 1.33 times Argon's tokens to finish a task, it wins on price today, and it can burn up to 2.67 times as many tokens after December 1 and still break even.
One workload, three price cards
A coding agent that reads 200M input tokens and writes 20M output tokens a month, with no cache hits:
| Model | Through Nov 30 | From Dec 1 | Arithmetic |
|---|
| Carbon or Argon | $600 | $1,200 | 200 x $2 + 20 x $10, then 200 x $4 + 20 x $20 |
| Hydrogen | $450 | $450 | 200 x $1.50 + 20 x $7.50 |
Add a 60% cache hit rate (120M cached input, 80M uncached) and the figures become $372 now and $744 from December 1 for Carbon or Argon, against $279 for Hydrogen. Cache hits cost one twentieth of the input price on all three.[3]
A worked bill for each model
These are illustrative workloads, not measurements; token counts differ by model, so each row uses its own profile.
| Model | Monthly workload | Through Nov 30 | From Dec 1 |
|---|
| Carbon | Nightly deep-review pipeline: 40M in, 12M out (thinking-heavy) | $200 | $400 |
| Argon | Interactive feature work: 150M in, 15M out | $450 | $900 |
| Hydrogen | Everyday agent fleet: 300M in, 25M out | $637.50 | $637.50 |
Carbon: 40 x $2 + 12 x $10 = $200, doubling to $400 at $4 and $20. Argon: 150 x $2 + 15 x $10 = $450, doubling to $900. Hydrogen: 300 x $1.50 + 25 x $7.50 = $637.50, or $381 at a 60% cache hit rate (120M uncached at $1.50, 180M cached at $0.075, output $187.50).
Carbon's expense is output tokens, so measure your output share before you commit a pipeline to it. Our AI model cost calculator lists Gemini 4 Argon, which carries the same rate as Carbon, so it covers the Carbon column. It does not list Hydrogen, so price Hydrogen with the per-token rates above.
Batch is the other lever. Google's pricing page advertises Batch API at a 50% cost reduction for paid projects, and all three models list Batch, Flex and Priority support.[3][4] The Batch, Flex and Priority rates for Carbon and Hydrogen are not confirmed here, so check them on the pricing page's Batch tab. The Interactions API does not support Batch yet, so batch jobs go through generateContent.[9]
How to call Carbon and Hydrogen
Google recommends the Interactions API for new projects, and the Gemini API docs use it for their Gemini 4 quickstarts.[9] The REST endpoint is POST https://generativelanguage.googleapis.com/v1beta/interactions with an x-goog-api-key header.[10] This curl call runs Carbon at its default thinking level. It sets max_output_tokens to 32,768 because non-streaming direct HTTP requests to Carbon cap there.[15] The jq filter reads the model_output step, matching the documented response shape.
curl -s -X POST https://generativelanguage.googleapis.com/v1beta/interactions \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-4-carbon",
"input": "Find the race condition in this retry queue and rewrite the lock acquisition.",
"generation_config": {
"thinking_level": "medium",
"max_output_tokens": 32768
}
}' | jq -r '.steps[] | select(.type=="model_output") | .content[] | select(.type=="text") | .text'
The Python SDK wraps the same call. Run pip install --upgrade "google-genai>=2.34.0" first: the Vertex AI docs require 2.34.0 or later for Gemini 4, while the Gemini API docs list 2.3.0 or later for the Interactions API.[15][9] The two floors differ, so install the higher one. The loop below sends one prompt to Hydrogen at high and Carbon at medium, which are each model's default, and prints token usage so you can price the run.
from google import genai
client = genai.Client() # credentials come from the environment, as in Google's quickstart
for model, level in [("gemini-4-hydrogen", "high"), ("gemini-4-carbon", "medium")]:
interaction = client.interactions.create(
model=model,
input="Find the race condition in this retry queue and rewrite the lock acquisition.",
generation_config={"thinking_level": level},
)
print(model, level)
print(interaction.output_text)
print("Usage:", interaction.usage)
Google pushes Carbon for planning and Hydrogen for execution: the Hydrogen docs say to route planning turns to gemini-4-carbon and execution turns to gemini-4-hydrogen in one session using previous_interaction_id.[7] Set generation_config on every call; it does not carry over.[9]
plan = client.interactions.create(
model="gemini-4-carbon",
input="Plan the migration of this service from callbacks to async/await. List ordered steps.",
generation_config={"thinking_level": "high"},
)
step_one = client.interactions.create(
model="gemini-4-hydrogen",
input="Implement step 1 of the plan and run the tests.",
previous_interaction_id=plan.id,
generation_config={"thinking_level": "medium"},
)
print(step_one.output_text)
On Vertex AI, which Google now brands Gemini Enterprise Agent Platform, the same model IDs work through generateContent on the global endpoint. This is Google's own curl example for Carbon.[15]
curl -X POST \
-H "Authorization: Bearer $(gcloud auth print-access-token)" \
-H "Content-Type: application/json" \
https://aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/global/publishers/google/models/gemini-4-carbon:generateContent \
-d '{
"contents": {
"role": "USER",
"parts": { "text": "How does AI work?" }
}
}'
Migration traps from Google's docs
Gemini 4 ignores temperature, top_p and top_k; on Hydrogen an out-of-range value returns a 400.[7] thinking_budget and the minimal thinking level return a 400, so use low, medium, high or, on Hydrogen only, xhigh.[15] Carbon can spend more of the output limit on thinking, so if your app caps max_output_tokens below 32K, Google says to raise it to at least 32,768. It also says to use background execution or set HTTP client timeouts to at least 15 minutes.[6] Non-streaming direct HTTP requests to Carbon and Argon cap at 32,768 output tokens, against 65,536 on Hydrogen.[15] For output past 64K, set background to true and raise max_output_tokens.[6]
Access: tiers, gates and regions
| Surface | Carbon | Hydrogen |
|---|
| Gemini API free tier | Not available | Free of charge, rate-limited; free-tier prompts may be used to improve Google's products |
| Gemini API paid tier | Yes, for all paying developers | Yes; paid-tier data is not used to improve products |
| Google AI subscriptions | AI Ultra: Antigravity, AI Studio, Gemini app, AI Mode | Paid AI Pro and Ultra in Antigravity; Gemini app model menu for everyone, Search for all users |
| Vertex AI / Gemini Enterprise Agent Platform | global endpoint, us and eu multi-regions | global endpoint, us and eu multi-regions |
| Cybersecurity variant | Carbon for cyber, via the Fairwind Program (allowlist) | Not applicable |
Table sources: launch posts, pricing page, Vertex quickstart.[1][2][3][15] Google says Hydrogen is rolling out globally.[2] Vertex lists Argon in two extra regions, europe-west2 and asia-southeast1.[15]
One gate hides in the rate-limit docs. Tier 1, which only needs a linked billing account, carries a spend limit of $10 per rolling 10 minutes, and Tier 2 raises it to $200 after $100 of payments and 3 days. Google says whether the limit applies depends on your billing history and account standing, and a breach returns a 429 RESOURCE_EXHAUSTED error.[11] If every dollar went to Carbon output at the $10 per 1M rate that runs through November 30, a Tier 1 project would hit that limit after roughly 1M output tokens in 10 minutes. Batch enqueued tokens at Tier 1 are 5,000,000 for Carbon and 3,000,000 for Hydrogen.[11]
The MKA benchmark: what it measures and why to be careful
Multimodal Knowledge Agent (MKA) is a benchmark Google DeepMind built with Artificial Analysis. It has 200 tasks across 10 verticals and 83 occupations. Each task is a multi-step project that Google says takes a seasoned professional three to eight hours, with a bundle of visual files such as drawings and schematics, and a rubric of roughly 70 yes/no checks weighted 75% on factual accuracy and 25% on visual design.[12] Agents run in the open-source Stirrup harness, and an LLM judge grades every deliverable; Google says domain experts agreed with it 96% of the time.[12]
Artificial Analysis publishes the leaderboard and runs the private 200-task set itself. It uses GPT-6 Luna (medium) as the judge and reports a headline score called Climb WSR, the weighted share of the harder rubric criteria satisfied. On the leaderboard as of November 16, 2026, Claude Fable 5.2 led at 57%, Gemini 4 Carbon (High) scored 54% and Claude Opus 5.5 scored 53%, across 21 models.[13] Google's headline is that Carbon ranks second;[1] the leaderboard framing is second place, 3 points behind first.
Caution is warranted for three reasons. A vendor co-designed the benchmark and then featured it in its own launch posts, which is a different thing from an independent benchmark that happens to include the vendor's model. The rubric and its 25% visual-design weight encode one definition of good work that may not match yours. And the Hydrogen claim does not reconcile cleanly: Google reports Hydrogen at 53.4% and says it outperforms other models, yet Hydrogen is not among the 21 models on the Artificial Analysis leaderboard as of November 16, 2026.[2][13] Treat 53.4% as vendor-reported until it shows up there.
In its favor: the test set is private, the judge is a non-Google model, and 10 public tasks are on Hugging Face so you can inspect the format.[12] For dense PDFs, drawings and spreadsheets, run the 10 public tasks or ten of your own documents through Hydrogen and Carbon and compare cost per passing task.
Which model for which workload
| Workload | Model | Reasoning |
|---|
| Chat, tool-using agents, PR iteration, live UI prototyping | Hydrogen, medium or high | Lowest price, fastest per Google's table, no end date listed for the rate |
| Parallel worker and verifier subagents | Hydrogen | Google positions it as the execution engine in multi-model setups |
| Hard bugs where Hydrogen at high or xhigh fails your tests | Argon | Frontier reasoning with fewer thinking tokens per turn than Carbon |
| Multi-package migrations, cross-subsystem root-cause hunts, architecture redesign | Carbon, as planner or coordinator | Built for extended verification; pair with Hydrogen for the edits |
| Dense multimodal synthesis: drawings, schematics, large PDF sets | Test Hydrogen first, then Carbon | MKA numbers are vendor-linked; compare cost per passing task on your files |
| Free prototyping, nothing sensitive | Hydrogen free tier | The only Gemini 4 model on the free tier, but prompts may train Google's products |
| High-volume classification and routing | Gemini 3.5 Flash-Lite | Google's Hydrogen guide keeps Flash-Lite for this, at $0.30 / $2.50 per 1M tokens[7][8] |
| Defensive security research | Carbon for cyber | Allowlist access through the Fairwind Program, not the public model |
If you already run Gemini 3.8 Flash, our Gemini 3.8 Flash guide covers what you are migrating from.
What is not confirmed
Four items are not confirmed by any published source: the Batch, Flex and Priority prices for Carbon and Hydrogen (the tabs exist on Google's pricing page, but the rates are not stated in the text this post relies on); any future Hydrogen price change (the page lists no end date, which is not a promise); the time zone of the November 30 cutover for Carbon and Argon; and Hydrogen's free-tier request quotas. Check them on the live pages before you sign off a budget.
Sources
Gemini 4 Carbon is now available, Google (published November 10, 2026; accessed November 16, 2026)
Introducing Gemini 4 Hydrogen, Google (published November 12, 2026; accessed November 16, 2026)
What's new in Gemini 4 Hydrogen (accessed November 16, 2026)
Multimodal Knowledge Agent (MKA) Benchmark, Artificial Analysis (accessed November 16, 2026)
15. Get started with Gemini 4, Google Cloud (accessed November 16, 2026)
Originally published at wowhow.cloud
Top comments (0)