
The Gemini 4 Carbon API is live on PoYo under the model ID gemini-4-carbon. You can call it through the OpenAI-compatible Chat Completions endpoint or Google's native generateContent format with one PoYo key, at $1.60 per 1M input tokens and $8.00 per 1M output tokens. Try it on the Gemini 4 Carbon model page before you write any code.
This overview covers what Google says the model is, the limits you have to design around, the price on PoYo and at Google, and which requests are worth routing to Carbon instead of Gemini 4 Argon.
What is Gemini 4 Carbon?
Google introduced Gemini 4 Carbon on October 28, 2026 as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities." It builds on Gemini 4 Argon, which launched two weeks earlier, and Google says it is "built for sustained agentic workflows": root-causing and patching bugs across large codebases, multi-agent coding, migrations, dense knowledge work, and scientific reasoning.
Access came in two steps:
| Date | What happened |
|---|---|
| October 28, 2026 | Google announces Gemini 4 Carbon; first access for a subset of cyber defenders through the Fairwind Program |
| November 10, 2026 | Available to all paying developers on the Gemini API paid tier (AI Studio) and to Gemini for business |
| November 16, 2026 | Callable on PoYo as gemini-4-carbon |
Google notes that some cybersecurity capabilities stay gated: developers apply to the Fairwind Program "to get access to Carbon for cyber on a rolling basis." Plan defensive-security features with that in mind.
Gemini 4 Carbon API at a glance
| Item | Value |
|---|---|
| PoYo model ID | gemini-4-carbon |
| Google model code | gemini-4-carbon (single stable version, no aliases listed) |
| Context window | 1,048,576 tokens, shared by input and output |
| Max output | 65,536 tokens by default (Google lists up to 1M only with "long decode") |
| Input types | Text, image, audio, video, PDF |
| Output type | Text |
| Knowledge cutoff | March 2026 |
| Free tier at Google | None |
| PoYo formats | Chat Completions and Gemini Native (not Responses or Anthropic Messages) |
Google's model page lists the 1M figure as a unified input-and-output limit, so a 1,000,000-token prompt leaves almost no room for the answer. PoYo's docs do not document a long-decode switch, so plan around 64K output per response on PoYo.
Endpoints and authentication
Both formats use Authorization: Bearer $POYO_API_KEY against https://api.poyo.ai. Responses come back in each format's native shape, without the code / data wrapper PoYo uses for media tasks.
| Format | Endpoint | Streaming | Max output field |
|---|---|---|---|
| OpenAI Chat Completions | POST https://api.poyo.ai/v1/chat/completions | "stream": true (SSE) | max_tokens |
| Gemini Native | POST https://api.poyo.ai/v1beta/models/gemini-4-carbon:generateContent | :streamGenerateContent (SSE) | generationConfig.maxOutputTokens |
Use Chat Completions when you already run the OpenAI SDK (set base_url to https://api.poyo.ai/v1). Use Gemini Native when you need inlineData parts for images and other files: PoYo's Chat Completions schema documents content as a plain string. Sending the model to an unsupported format, such as /v1/responses, returns HTTP 400 with the list of supported URIs. Request shapes are in the Gemini Native docs.
Google also lists function calling, structured outputs, thinking levels, and Google Search grounding for this model. PoYo's published schemas for these two endpoints document messages, contents / parts, sampling fields, generationConfig, and safetySettings only, so check the docs before you depend on anything else.
Gemini 4 Carbon API pricing: PoYo vs Google
PoYo bills per token in credits (1 credit = $0.005):
| Meter (per 1M tokens) | PoYo price | PoYo credits | Google, through Nov 30, 2026 | Google, from Dec 1, 2026 |
|---|---|---|---|---|
| Input | $1.60 | 320 | $2.00 | $4.00 |
| Output (incl. thinking tokens) | $8.00 | 1,600 | $10.00 | $20.00 |
| Cache read | $0.08 | 16 | $0.10 | $0.20 |
PoYo's rates are 20% below Google's introductory input and output prices. Google's pricing page says thinking tokens are included in the output rate, and that Carbon supports implicit context caching only. The Google prices above are standard tier. Google also publishes batch rates, which are not part of this PoYo endpoint. PoYo prices are as listed on November 16, 2026; the model page is the current source.
A 200K-token repository review with an 8K-token answer costs 0.2 × $1.60 + 0.008 × $8.00 = $0.384 (76.8 credits) on PoYo. Full worked examples are in Gemini 4 Carbon API pricing.
Which requests to route to the Gemini 4 Carbon API
Benchmarks only matter here as routing evidence. From Google's published table:
| Workload | Benchmark | Carbon | Argon | Route to Carbon? |
|---|---|---|---|---|
| Terminal-based agentic coding | Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | Yes, for long multi-step tasks |
| Long-horizon agentic coding | SWE-sweep | 9.6% | 4.3% | Yes |
| Retrieval across 256K–1M tokens | GraphWalks (BFS F1) | 96.2% | 84.2% | Yes, when the prompt is huge |
| Business automation across tools | AutomationBench (Zapier) | 53.9% | 51.3% | Small gap; test both |
| Vulnerability finding | SEC-bench Pro | 88.2% | 83.4% | Yes, if you have Fairwind access |
| Patch quality for weaknesses | CWE-bench | 71.0% | 68.0% | Yes |
The same table has GPT-6 Astra ahead on FrontierSWE 2 (65.5% vs 61.2%) and Claude Opus 5.5 ahead on OSWorld-2.1 (87.2% vs 81.5%). For a three-way breakdown, see Gemini 4 Carbon API vs Claude Opus 5.5 vs GPT-6 Astra.
Carbon vs Argon: one routing rule
Carbon and Argon have the same input and output price, on PoYo and at Google. The difference is tokens per task: Google says Carbon works by "using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops," and that "Argon remains the ideal choice for the majority of software engineering and knowledge work." So send everyday coding and document work to gemini-4-argon. Send the hardest debugging, migrations, million-token retrieval, and security work to gemini-4-carbon. Because both use the same request shape, the switch is just the model ID. Details are in Gemini 4 Carbon API vs Gemini 4 Argon and the Gemini 4 Argon API guide.
Gemini 4 Carbon API quick start
curl https://api.poyo.ai/v1/chat/completions \
-H "Authorization: Bearer $POYO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-4-carbon",
"messages": [
{"role": "system", "content": "You are a senior engineer. Answer with a numbered plan."},
{"role": "user", "content": "Plan a migration of a Flask service from SQLAlchemy 1.4 to 2.0."}
],
"max_tokens": 8192
}'
The answer is in choices[0].message.content and token counts are in usage. The same request through the OpenAI SDK, streaming, image input, and the native format are in the Gemini 4 Carbon API guide. For agent loops and long-context coding, see Gemini 4 Carbon API for coding agents.
Frequently asked questions
What is the Gemini 4 Carbon API model ID?
Use gemini-4-carbon. Pass it in "model" for /v1/chat/completions, or in the URL path for /v1beta/models/gemini-4-carbon:generateContent.
How much does the Gemini 4 Carbon API cost?
On PoYo: $1.60 input, $8.00 output, and $0.08 cache read per 1M tokens (320 / 1,600 / 16 credits). Google charges $2.00 / $10.00 / $0.10 through November 30, 2026 and $4.00 / $20.00 / $0.20 from December 1, 2026.
Is the Gemini 4 Carbon API compatible with the OpenAI SDK?
Yes. Set base_url="https://api.poyo.ai/v1", use your PoYo key, and set model="gemini-4-carbon" in chat.completions.create.
What is the Gemini 4 Carbon context window and max output?
1,048,576 tokens shared between input and output, with 65,536 output tokens by default. Inputs can be text, images, audio, video, and PDF; the output is text.
Is there a free tier for the Gemini 4 Carbon API?
Google's pricing page says Gemini 4 Carbon is not available on the free tier. PoYo is pay-as-you-go with credits.
Should I use the Gemini 4 Carbon API or the Gemini 4 Argon API?
Same per-token price; Carbon spends more tokens per task. Google recommends Argon for most software engineering and knowledge work, and Carbon for the hardest long-horizon, long-context, and security tasks.