Gemini 4 Carbon API

Gemini 4 Carbon at a glance

Release

Announced 2026-10-28, first for cyber defenders in the Fairwind Program; generally available from 2026-11-10 on the paid Gemini API tier, in AI Studio, Android Studio and Gemini for business, and for Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search

Positioning

Google’s strongest reasoning model, with frontier performance in cybersecurity and long-horizon agentic workflows, meant for the hardest tasks. Google still recommends Argon for most software engineering and knowledge work; Carbon spends more tokens on deeper exploration and verification

Context and output

1,048,576-token context window shared by input, thinking and output; up to 65,536 output tokens per synchronous request, about 1M with long decode (background mode). Accepts text, images, audio, video and PDF and outputs text; knowledge cutoff March 2026

Thinking levels

Three levels, low / medium / high, with medium as the API default and Google’s recommendation for most tasks; thinking tokens are billed at the output rate, and max_output_tokens caps thinking and output together

Official pricing

Same as Argon. Through 2026-11-30: $2 input / $10 output per million tokens, $0.10 cache read; from 2026-12-01 everything doubles to $4 / $20 / $0.20. Batch and Flex are half price and Priority is 1.8x; no long-context surcharge; implicit caching only (prefixes of at least 4,096 tokens); no free API tier

API features

Function calling, structured outputs, code execution, Google Search and Maps grounding, URL context, file search, computer use, Batch, Flex and Priority; no Live API, image or audio generation, or explicit caching. temperature, top_p and top_k are ignored

Subscriptions

In Antigravity only Google AI Ultra and Enterprise can use Carbon; AI Pro, Plus and free plans cannot (Argon is open to Pro). Antigravity also launched /swarm multi-agent orchestration in preview

Model ID

gemini-4-carbon (the only stable version so far)

Checked 2026-11-11 · Sources: Model card · Gemini API · Model · Gemini API · What’s new · Gemini API · Pricing · Google · Antigravity · Artificial Analysis

Official benchmarks: Gemini 4 Carbon vs Argon vs GPT-6 Astra vs Claude Opus 5.5 vs Claude Fable 5.1

From Google DeepMind’s Gemini 4 Carbon model card (as of October 2026). These are vendor-run results, and Google also supplied the competitor scores; the two rows marked AA were run by Artificial Analysis; “—” means Google listed no score.

In Google’s table Carbon leads 15 of 17 rows, trailing only on FrontierSWE 2 (GPT-6 Astra 65.5%) and OSWorld 2.1 (Opus 5.5 87.2%). Its largest gains over Argon are long context (GraphWalks 256K–1M 96.2% vs 84.2%), ExploitGym (54.8% vs 42.0%), Terminal-Bench Science and Terminal-Bench 4.0. Two rows are not shown above: AA Index v5.0 is 59 for Carbon, 50 for Argon, 57 for Astra, 58 for Opus 5.5 and 49 for Fable 5.1; on Vending-Bench 2 Carbon ends with the highest balance, $16,484. Claude Fable 5.2 is not in Google’s table. In Anthropic’s Fable 5.2 launch table Fable 5.2 beats Carbon on FrontierSWE v2 (69.9% vs 61.2%), OSWorld 2.1 (89.3% vs 81.5%) and Terminal-Bench Science (74.6% vs 64.8%), while Carbon is higher on AutomationBench (53.9% vs 51.4%).

Which to use: Gemini 4 Carbon, Gemini 4 Argon, Claude Opus 5.5 or Fable 5.2

Versus Gemini 4 Argon

Both cost exactly the same, and switching only means changing the model ID; the request format does not change. Carbon beats Argon on every row of the model card, and on the Artificial Analysis Intelligence Index (v5.0.1, high) Carbon scores 60 to Argon’s 50. Carbon used 205M output tokens to run the whole index, about 1.8x Argon’s 113M, yet costs $7.76 per task at $4 / $20, below Argon’s $9.95, at 25.6 minutes per task on average. The cost is latency: about 67 seconds to Carbon’s first answer token against about 38 for Argon. Google’s advice: use Argon for everyday coding and knowledge work, where it answers faster with fewer tokens per turn, and Carbon for the hardest problems, long-horizon agents, multi-agent work and security audits.

Versus Claude Opus 5.5

From December 1 both list at $4 / $20. In Google’s model card Carbon leads Opus 5.5 on Terminal-Bench 4.0 (65.2% vs 59.6%), AutomationBench, long context and cybersecurity, while Opus 5.5 leads on OSWorld 2.1 (87.2% vs 81.5%) and FrontierSWE 2. On the Artificial Analysis Intelligence Index Carbon scores 60 at high for $7.76 per task, and Opus 5.5 scores 59 at max for $10.04, with about 370 seconds to its first answer token. For computer-use agents start with Opus 5.5; for reading hundreds of thousands of tokens of code or documents at once, or for security audits, try Carbon first. Opus 5.5 lets you place cache breakpoints with cache_control, while Carbon only has implicit caching.

Versus Claude Fable 5.2 and GPT-6 Astra

Claude Fable 5.2 lists at $8 / $40, twice Carbon’s December price. It tops the Artificial Analysis Intelligence Index at 65 (max), but costs $26.54 per task, 3.4x Carbon, and in Anthropic’s table it is clearly ahead on FrontierSWE v2 and OSWorld. GPT-6 Astra lists at $10 / $50 and scores 57 at max for $6.24 per task. Choose Fable 5.2 for the highest scores when budget allows; at Carbon’s price point, choose Carbon for long context and cost.

Setting the thinking level and timeouts

Carbon has only low, medium and high. Google recommends medium for most tasks, including complex code and agents, and says to pick the lowest level that meets your quality bar. High has the slowest first token: Artificial Analysis measured about 67 seconds to the first answer token and 25.6 minutes per task. Google recommends client timeouts of at least 15 minutes, or background execution, and raising any max_tokens below 32K to at least 32,768, because thinking tokens count toward that cap. Through Wokey, stream the response, choose the level with reasoning_effort, and expect medium when you omit it.

Official migration notes

Independent evaluations and official figures

• Artificial Analysis: On Intelligence Index v5.0.1 Carbon scores 60 at high, 4th of 74 models, against a median of 40. Running the full index took 205M output tokens (median 72.8M), which puts it among the most verbose models, or $7.76 per task at $4 / $20. Output speed is 80 tokens per second, with 66.85 seconds to the first answer token and 25.6 minutes per task on average. Artificial Analysis sums it up as “amongst the leading models in intelligence, but somewhat expensive” and “notably slow per task”. Every Wokey meter is 30% of Google’s December list price, so the same task comes to about $2.33.
• Google launch post: Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at more than 2x lower cost per task than the next best external model, and second on Artificial Analysis’ new Multimodal Knowledge Agent benchmark at more than 6x lower average cost per task. On safety, its success rate on Gray Swan’s cyberattack requests is 2.17% (Argon 9.86%), indirect prompt injection succeeded 0% of the time, and it passes 92% of 375 honesty tasks (Argon 70%). These are Google’s own figures.
• Cursor: Cursor added Carbon on launch day and says that at high effort it is the top model on CursorBench, at 60.4%.
• Google Antigravity: Antigravity launched /swarm in preview the same day: a coordinator splits the task across parallel workers that each edit code in their own git worktree, plus adversarial verifiers that run tests and challenge diffs. Google suggests Carbon for the coordinator, architecture, security-audit and verifier roles and lighter models for routine edits, and warns that parallel agents use quota much faster than a normal chat. Only Ultra and Enterprise can pick Carbon in Antigravity.

Calling Gemini 4 Carbon through Wokey: parameters and limits

This is how the Wokey gateway handles gemini-4-carbon. Coming from Argon, check timeouts and max_tokens besides the model ID.

• The interface is OpenAI-compatible /v1/chat/completions: set the base URL to https://api.wokey.ai/v1, use your Wokey API key, and set model to gemini-4-carbon. opencode, OpenClaw, Hermes Agent and other OpenAI-compatible clients are configured the same way.
• reasoning_effort values low, medium and high map directly to Carbon’s three thinking levels; without it, or with any other value such as minimal, xhigh or max, the request runs at medium (the same default as Google’s API).
• Without max_tokens the limit defaults to 65,536, which is also Wokey’s current output ceiling; long decode and background mode are not supported yet. Thinking tokens count toward max_tokens and are billed at the output rate, so do not set it too low.
• Carbon thinks for a long time, and at high the first answer token often takes over a minute. Use stream: true and set client timeouts to 15 minutes or more.
• Google ignores temperature, top_p and top_k, so you can leave them out.
• Text and image input are accepted today; audio, video and PDF input are not yet open on Wokey.
• Tool calling: the gateway stores the thought signatures Carbon returns, so send back tool_calls and tool results as usual on the next turn without handling signatures yourself.
• Structured outputs: use response_format with json_schema, and keep the schema within the OpenAPI subset Gemini accepts (type, properties, required, items, enum, description, and nullable via anyOf with null). json_object, or a schema with $ref, oneOf or pattern, currently returns 400 structured_output_supply_unavailable.
• Caching is Google’s implicit caching, with no cache_control; keep long sessions on a stable prefix to hit it. Cache reads cost 5% of the input rate.

Calling Gemini 4 Carbon through Wokey (/v1/chat/completions)

curl https://api.wokey.ai/v1/chat/completions \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-4-carbon",
    "reasoning_effort": "medium",
    "stream": true,
    "max_tokens": 65536,
    "messages": [{ "role": "user", "content": "Audit this module for injection bugs and propose patches" }]
  }'

Catalog pricing snapshot

Prices are per 1M tokens; request-time pricing controls settlement.

Official price source: https://aistudio.google.com/docs/pricing

Model capabilities

Context

1,048,576 tokens

Max output

65,536 tokens

Streaming

Supported

Tools

Supported

Vision

Supported

Supported API forms

/v1/chat/completions

Send a request

/v1/chat/completions

curl https://api.wokey.ai/v1/chat/completions \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gemini-4-carbon","messages":[{"role":"user","content":"Hello"}]}'

Use Gemini 4 Carbon in your tools

The gateway is https://api.wokey.ai (OpenAI-compatible clients usually take https://api.wokey.ai/v1), and the model ID is gemini-4-carbon. Open the guide for the client you use.

• Use Gemini 4 Carbon in opencode: OpenAI Chat Completions · base URL and API key setup
• Use Gemini 4 Carbon in DeepSeek Harness: OpenAI Chat Completions · base URL and API key setup
• Use Gemini 4 Carbon in Cherry Studio: OpenAI Chat Completions · base URL and API key setup
• Use Gemini 4 Carbon in WorkBuddy: OpenAI Chat Completions · base URL and API key setup
• Use Gemini 4 Carbon in OpenClaw: OpenAI Chat Completions · base URL and API key setup
• Use Gemini 4 Carbon in Hermes Agent: OpenAI Chat Completions · base URL and API key setup
• Use Gemini 4 Carbon in Magpie: OpenAI Chat Completions · base URL and API key setup
• OpenRouter alternative: Wokey vs OpenRouter: Compare token prices, payment fees, API forms, and upstream sources.
• Verifiable AI API: which models and API forms carry response proofs: Response proofs currently cover only raw passthrough Claude Messages and GPT Responses on official routes.

Frequently asked questions

Related models and documentation

• Gemini 3.5 Flash
• Gemini 4 Argon
• Gemini 3.8 Flash
• API documentation
• All models and pricing
• Provider Node