How to use Gemini 4 Carbon: access, requests and limits

Gemini 4 Carbon became available to paying developers on November 10, 2026, two weeks after Google first released it to cybersecurity defenders. Its request model ID is gemini-4-carbon, it accepts text, images, video, audio and PDFs, and it returns text.^[1][2] Many guides written in late October still describe it as unavailable, so check the date on anything you copy.

This guide covers how to use Gemini 4 Carbon today: where you can run it, how to write a first request in the native Gemini format, and the output limits that catch people moving from older Gemini models. reAPI does not provide inference for this model. The Gemini 4 Carbon overview on reAPI includes a request editor, and the service itself runs on SeedRouter.

TL;DR

• Developers get Gemini 4 Carbon through a paid Gemini API tier; it is not on the free tier. Google AI Ultra subscribers can also use it in the Gemini app, Google Antigravity and AI Mode in Search.^[1][4]
• Moving from Gemini 4 Argon needs only a model ID change, but Carbon spends more thinking tokens per turn, so leave room in the output budget and the client timeout.^[5]
• thinkingLevel is the main control: low, medium (default) or high. Temperature, top-p and top-k are ignored.^[5]
• A non-streaming REST call returns at most 32,768 output tokens per hop. Asking for more without a continuation token returns a 400 error.^[6]
• On SeedRouter, Gemini 4 Carbon uses the native Gemini format, with generateContent and streamGenerateContent.^[3]

Where you can use Gemini 4 Carbon today

Google's November 10 announcement lists several separate routes. They are different products with different accounts, so an Ultra subscription in the Gemini app is not an API key, and a paid API key does not unlock the consumer app.^[1]

Free access is the most common question, and the answer is no for the API. Google's pricing page marks every Gemini 4 Carbon line as "Not available" on the free tier.^[4] The consumer route requires Ultra, not the regular paid plan.

Carbon or Argon for this task

Google positions Argon as the right choice for most software engineering and knowledge work. Carbon is meant for tasks that strain a model's reasoning, such as multi-step debugging across a large codebase or open-ended research, and it uses more tokens than Argon to explore and verify its work.^[1] Google's migration notes say the same thing in practical terms: keep Argon where you want faster responses and lower token spend per turn.^[5]

The two models share a request schema, so you can route a hard task to gemini-4-carbon and keep everyday traffic on gemini-4-argon without maintaining two clients. Google also launched Carbon at the same introductory token price as Argon.^[1] Since Carbon thinks longer, the same prompt can still cost more on Carbon even at an identical rate per token. Thinking tokens are billed as output.^[4]

On the Gemini API, Google lists $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1.^[4] Cached input is $0.10 per million tokens during the introductory period and $0.20 afterward. These are Google's rates for its own API. Other services set their own prices, so check SeedRouter's model page for what you would actually pay there.

Prepare your first Gemini 4 Carbon request

Pick a task whose answer you can check, such as a short function with a known bug. Starting with a whole repository makes it hard to tell an access problem from a request problem or a weak prompt.

In the native Gemini format the model goes in the URL path, not the body, and contents is the only required field.^[3] Here is a body for POST /v1beta/models/gemini-4-carbon:generateContent:

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "This retry loop re-reads an account balance before taking the lock. Explain the race condition and propose a fix with one test."
        }
      ]
    }
  ],
  "generationConfig": {
    "thinkingConfig": { "thinkingLevel": "medium" },
    "maxOutputTokens": 32768
  }
}

The 32768 budget follows Google's advice to set at least 32,768 when you set the field explicitly, because Carbon can spend much of the limit on thinking.^[5] It is also the ceiling for one non-streaming REST call, which the next section explains.

SeedRouter documents this same path on its own host, api.seedrouter.ai, with a bearer token in the Authorization header. Google's own REST examples use an x-goog-api-key header instead.^[3][6] Keep each key with the service that issued it. A Google key will not authenticate on SeedRouter, and a SeedRouter key will not work against Google's host.

Leave out what Gemini 4 Carbon ignores or rejects. Google says temperature, topP and topK are ignored if you send them, and explicit context caching is not supported. Implicit caching applies automatically to prompts of at least 4,096 tokens, so put large static material such as a codebase or long document at the start of the prompt.^[5] If your client still sends thinkingBudget from an older model, replace it with thinkingLevel; Google's Agent Platform guide says thinking_budget and MINIMAL return HTTP 400 on Gemini 4 models there.^[7]

Streaming, output caps and continuation

Gemini 4 Carbon has a unified window of 1,048,576 tokens that input, thinking and output share. A single synchronous call stops far below that.^[2] Google's long-decode guide sets the per-hop caps:^[6]

Continuation works hop by hop. When a hop reaches its cap and budget remains, the response ends with finishReason: "CONTINUATION" and a continuationToken. Send the identical request again with that token added, and do not append the partial answer to contents, since the token already carries the model's state. Keep going while the finish reason is CONTINUATION.^[6]

A hop can return empty text. On hard problems the model may spend a whole hop on thinking, and Google treats an empty text with a continuation token as normal progress rather than a failure.^[6]

Plan for slow responses too. Google recommends HTTP read timeouts of 20 minutes for long decode, and SeedRouter warns that a non-streaming request running longer than about 100 seconds can end with HTTP 524 before the answer arrives. For long tasks, SeedRouter recommends streamGenerateContent?alt=sse.^[3][6] A dropped connection does not prove that generation failed, so check the partial result before sending the same request again.

Choosing a thinking level

thinkingLevel trades quality against latency and token use.^[5]

Google's advice is to run your own evaluation at all three levels and pick the lowest one that meets your quality bar.^[5] A level that worked on Argon is a reasonable starting point but not a conclusion, because Carbon spends tokens differently.

Check the response before scaling up

Read more than the answer text. A native response carries candidates[].content.parts, finishReason and usageMetadata, and SeedRouter's documentation also lists continuationToken, safetyRatings and promptFeedback among the fields worth reading.^[3]

• finishReason: "MAX_TOKENS" means the output budget ran out. Google says you can resend with the returned token and a higher maxOutputTokens, or start over and pay for both runs.^[6]
• thoughtsTokenCount shows the thinking tokens you were billed for as output. A short visible answer can sit on top of a large thinking count.^[4][6]
• REST returns usage for one hop at a time. Add the fields across hops for cost, and take the prompt size from the first hop only.^[6]
• Streamed usage on SeedRouter is cumulative, so read the latest counters instead of summing every chunk.^[3]

Keep the complete content parts, including thoughtSignature, when you build conversation history.^[3] For a code fix, run the tests before you accept the patch, since a 200 response only shows that the service accepted the request.

What Google reported at launch

Google's October 28 announcement put Gemini 4 Carbon first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. It reported 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, all evaluated without guardrails.^[8] These are vendor-reported results. Test the model on your own tasks before you route production traffic to it.

FAQ

Is Gemini 4 Carbon free?

Not on the Gemini API. Google's pricing page lists it as unavailable on the free tier, and consumer access in the Gemini app requires a Google AI Ultra subscription.^[1][4] Researchers can apply for six months of free Antigravity access, subject to selection.

Is Gemini 4 Carbon available now?

Yes, since November 10, 2026, for paying Gemini API developers, Gemini for business and Google AI Ultra subscribers. Before that it was limited to defenders in the Fairwind Program.^[1]

How much does Gemini 4 Carbon cost?

On Google's API, $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20.^[4] Thinking tokens count as output. SeedRouter publishes its own rates on its model page.

What is the Gemini 4 Carbon context window?

1,048,576 tokens shared by input, thinking and output. The default synchronous output limit is 65,536 tokens, or 32,768 for a non-streaming REST call, and long decode can go further.^[2][6]

Can I switch from Gemini 4 Argon to Carbon?

Yes. Google says the upgrade needs no request schema change, only the model ID. Allow for longer thinking in your output budget and timeouts.^[5]

Can I use Gemini 4 Carbon on reAPI?

No. reAPI does not provide inference, billing or keys for this model. The overview page helps you prepare a request, and SeedRouter is where you run it.

Getting a first Gemini 4 Carbon run right

Start with a small task, use gemini-4-carbon in the URL, set thinkingLevel and maxOutputTokens deliberately, and use streaming when a task may run long. Read the finish reason and token counts before making the prompt bigger. When you are ready, open Gemini 4 Carbon on SeedRouter and check access for your account.

References

1. Google. Gemini 4 Carbon is now available. Published November 10, 2026, retrieved November 11, 2026 from blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available.
2. Google AI for Developers. Gemini 4 Carbon model page. Retrieved November 11, 2026 from ai.google.dev/gemini-api/docs/models/gemini-4-carbon.
3. SeedRouter. Gemini 4 Carbon API documentation. Retrieved November 11, 2026 from seedrouter.ai/docs/gemini-4-carbon. Model page: seedrouter.ai/models/gemini-4-carbon. Live model execution was not tested for this guide.
4. Google AI for Developers. Gemini Developer API pricing: Gemini 4 Carbon. Retrieved November 11, 2026 from ai.google.dev/gemini-api/docs/pricing.
5. Google AI for Developers. What's new in Gemini 4 Carbon. Retrieved November 11, 2026 from ai.google.dev/gemini-api/docs/whats-new-gemini-4-carbon.
6. Google AI for Developers. Long decode (generateContent). Retrieved November 11, 2026 from ai.google.dev/gemini-api/docs/long-decode.
7. Google Cloud. Get started with Gemini 4. Gemini Enterprise Agent Platform documentation, retrieved November 11, 2026 from docs.cloud.google.com/gemini-enterprise-agent-platform/models/start/get-started-with-gemini-4.
8. Google. Gemini 4 Carbon: our next frontier model. Published October 28, 2026, retrieved November 11, 2026 from blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon.