Gemini · Model guide
Gemini 4 Carbon
Gemini 4 Carbon is Google’s most capable reasoning model, built on Gemini 4 Argon for long-horizon agentic coding, high-precision multimodal knowledge work, and scientific reasoning. It spends more thinking per turn than Argon to explore and verify its work, so responses take longer. It accepts text, images, audio, and video and returns text within a shared 1M-token input and output window.
What this model is good at
What the vendor positions it for, and which of our models to reach for instead.
VendorGoogle positions Carbon for tasks that test the limits of reasoning depth: complex multi-step coding and debugging, multimodal knowledge work, and open-ended scientific research.source
Deep debugging and agentic coding · Google recommends it for complex root-cause debugging and architectural design, running build-and-verify loops across complex repositories.(vendor claim)source
High-precision knowledge work · Google highlights stronger source grounding with fewer unsupported claims across dense PDFs, spreadsheets, and slides, with strict output-schema adherence.(vendor claim)source
Mathematical and scientific reasoning · Google lists formal proofs, experimental data analysis, and reasoning across physics, math, and life sciences among its intended uses.(vendor claim)source
Open-ended research · Google positions it for open-ended scientific research and discovery that needs thorough hypothesis exploration and extended verification.(vendor claim)source
- gemini/gemini-4-argonfaster —Choose Gemini 4 Argon for interactive coding, day-to-day feature work, and fast production loops, where Google says it responds faster with fewer thinking tokens per turn.(vendor claim)source
- gemini/gemini-3.1-pro-previewpredecessor —Keep Gemini 3.1 Pro Preview when an existing Gemini 3 integration has already been validated and you want to compare it with Carbon before migrating.source
Pricing and billing
Billed per token. Cached input is charged at the cache-read rate.
| Price / 1M tokens | You pay |
|---|---|
| Input | $ 4.00$ 2.00 |
| Output | $ 20.00$ 10.00 |
| Cache read | $ 0.20$ 0.10 |
| Cache write | $ 4.00$ 2.00 |
+ 2.0K × $ 10.00
Capabilities and limits
What CrossModel guarantees across every route this model can take right now.
- Context window
- 1.0M tokens
- Max output
- 65.5K tokens
- Input / output modalities
- Text + Image + Audio + Video → Text
- Streaming
- Supported
- Tool calling
- Supported
- Structured output (JSON)
- Supported
- Image input
- Supported
- Thinking mode
- Supported· always on
- Reasoning effort
- low · medium · high
- Available endpoints
- /v1/chat/completions · /v1/responses · /v1/messages
Published benchmarks
Scores the vendor reported, with the evaluation setup each one came from.
- Agents' Last Exam v1.1score·Google DeepMind · 2026-11-10effort highALE-Clawsetup not fully disclosed
Google self-computed averaged rubric score on the 122-task v1.1 set (task version of 2026-10-17, without the 2026-10-26 update); default ALE-Claw harness, five-hour window, safety filters on (flagged responses return empty strings and the episode continues). Not comparable with the binary pass rate. Same table: Gemini 4 Argon 69.0%, GPT-6 Astra 73.1%, Claude Opus 5.5 71.6% (both from the public leaderboard). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-carbon
77.4%
- SWE-sweepscore·Google DeepMind · 2026-11-10effort highmini-swe-agentsetup not fully disclosed
Google self-computed with a mini-swe-agent harness; reports the subtask-weighted reward. Highest thinking setting, single attempt. Same table: Gemini 4 Argon 4.3%, GPT-6 Astra 9.5%, Claude Opus 5.5 3.6%. Competitor figures in that table come from the public SWE-sweep leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-carbon
9.6%
- OSWorld 2.1 (offline subset, partial score)score·Google DeepMind · 2026-11-10effort highGemini CUA3 runsCompaction following Claude Opus 5.5 system cardsetup not fully disclosed
Google self-computed partial score on the 82-task offline (no-internet) subset, maximum over three single-attempt runs. Official OSWorld docker and evaluator, 1080p, max 500 steps; updated Gemini CUA harness with parallel batch tool calls, all screenshots kept in history, PyAutoGUI actuation with screenshot-only observation, safety filters on. Same table: Gemini 4 Argon 74.7%, Claude Opus 5.5 87.2%, Claude Fable 5.1 76.8% (from their own published evaluations). Methodology: https://deepmind.google/models/evals-methodology/gemini-4-carbon
81.5%
- GraphWalks BFS 256K to 1Mf1·Google DeepMind · 2026-11-10effort highsetup not fully disclosed
Google self-computed F1 on the 200 problems with context lengths from 256K to 1M tokens. Same table: Gemini 4 Argon 84.2% (matches its own model card), GPT-6 Astra 71.8%, Claude Opus 5.5 66.8%. Full reproduction settings not disclosed. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-carbon
96.2%
Vendor-reported numbers, not CrossModel measurements. Scores are only comparable when the benchmark version, metric and evaluation setup match, so nothing here is averaged or ranked.
Use it in your tools
Point the base URL at CrossModel and paste this model ID — every tool below has a setup guide.
Frequently asked questions
What is Gemini 4 Carbon?
How much does Gemini 4 Carbon cost?
Does Gemini 4 Carbon support tool calling and structured output?
Which endpoint do I call?
Can I try it without writing code?
Chat first, integrate later
Chat first, then wire it in once you like the answers.