Gemini 4 Carbon is Google's AI model, released 28/10/2026. Gemini 4 Carbon is Google DeepMind's top Gemini 4 model, based on Gemini 4 Argon and aimed at long-horizon agentic work, cybersecurity and scientific reasoning. The vendor's table has it ahead of Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on most rows, with Astra ahead on FrontierSWE 2 and Opus 5.5 ahead on OSWorld 2.1. Cybersecurity results reflect access through the Fairwind Program for verified defenders.
Visit Gemini 4 CarbonAI-Powered
Leverages advanced AI technology to deliver cutting-edge capabilities and results.
Fast & Efficient
Optimized performance ensures quick results without compromising on quality.
Purpose-Built
Specifically designed for llms tasks and workflows.
How does Gemini 4 Carbon score on independent benchmarks?
Gemini 4 Carbon ranks #4 of 28 models on the Artificial Analysis Intelligence Index (v5.0.1) with a score of 60, as measured on 22 November 2026.
Where does Gemini 4 Carbon rank against other models?
How good is Gemini 4 Carbon at agentic coding?
How does Gemini 4 Carbon do on individual tests?
| Evaluation (run by Artificial Analysis) | Gemini 4 Carbon | Best of any model |
|---|---|---|
| Terminal-Bench 4.0 | 65.2% | 67.2% |
| FrontierCode 2.0 | 40.6% | 48.8% |
| Humanity's Last Exam | n/a | 74.3% |
| GDPval-AA (Elo) | 1733 | 1866 |
How has Gemini 4 Carbon's score changed over time?
We have tracked Gemini 4 Carbon on Intelligence Index v5.0.1 since 22/11/2026. A point is added each day the score, coding index or price changes, so this chart fills in over time.
What does the Gemini 4 Carbon system card report?
Gemini 4 Carbon is Google DeepMind's top Gemini 4 model, based on Gemini 4 Argon and aimed at long-horizon agentic work, cybersecurity and scientific reasoning. The vendor's table has it ahead of Argon, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on most rows, with Astra ahead on FrontierSWE 2 and Opus 5.5 ahead on OSWorld 2.1. Cybersecurity results reflect access through the Fairwind Program for verified defenders.
These are Google DeepMind's own published results (self-reported, with the vendor's harness and settings), so they are not directly comparable with other labs' numbers. Sources: Gemini 4 Carbon Model Card (undefined/11/2026), Gemini 4 Carbon Model Evaluation (approach, methodology and results) (undefined/10/2026), Introducing Gemini 4 Carbon (28/10/2026). Last checked 22/11/2026.
Other
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index v5.0with tools; highest thinking | 59 points | 50 | 57 | 58 |
| Vals Index v2.2Vals AI sourced | 70.4% | 69.2% | 65% | 70.2% |
| CWE-bench v1 (cybersecurity)programmatic pass@1; accessed through the Fairwind Program | 71% | 68% | 68% | 67% |
| ExploitGym (cybersecurity)Antigravity harness, mitigations off, single attempt; 6 hour default and 10 hour extended timeouts reported in methodology; Fairwind Program | 54.8% | 42% | 42.4% | 34.5% |
| SEC-bench Pro (cybersecurity)pass@1 on 183-task May 2026 JavaScript engine split, 3 attempts, Antigravity harness; Fairwind Program | 88.2% | 83.4% | 85.4% | n/a |
| WebDev Arenarank #4, up 89 points per launch post | 1768 Elo | n/a | n/a | n/a |
Coding
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| SWE-sweepsubtask-weighted reward; mini-swe-agent harness; self-computed | 9.6% | 4.3% | 9.5% | 3.6% |
| FrontierSWE 2Proximal official public leaderboard | 61.2% | 55% | 65.5% | 62.3% |
| Terminal-Bench 4.0 (AA)Artificial Analysis sourced; cheating reduced to 0.0% per launch post | 65.2% | 57.1% | 59.1% | 59.6% |
Agents & tool use
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Agents' Last Exam v1.1score; 122-task set, ALE-Claw harness, 5-hour window, safety filters on; self-computed | 77.4% | 69% | 73.1% | 71.6% |
| OSWorld 2.182-task offline subset, max of 3 runs, partial score; self-computed | 81.5% | 74.7% | 80.9% | 87.2% |
| AutomationBenchZapier private set, official leaderboard | 53.9% | 51.3% | 41.4% | 42.5% |
| Vending-Bench 2final bank balance; Andon Labs leaderboard | 16484 USD | 13718 | 15515 | 9235 |
Multimodal
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Chartographyno tools; Surge leaderboard | 72.5% | 71.6% | 71% | 66.3% |
| Blueprint-Bench 2similarity score; Andon Labs leaderboard | 58.6% | 54.4% | 49.7% | 51.2% |
Long context
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| GraphWalks (256k to 1M)BFS F1 on 200 problems between 256k and 1M tokens; self-computed | 96.2% | 84.2% | 71.8% | 66.8% |
Reasoning & maths
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| RiemannBenchSurge public leaderboard | 80% | 76% | 72% | 69.6% |
Knowledge & science
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Terminal-Bench Science 0.1 (AA)Artificial Analysis sourced | 64.8% | 52.9% | 63.3% | 59% |
Safety & honesty
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Gray Swan IPI (attack success rate, lower is better)indirect prompt injection; per launch post | 0% | 0.7% | n/a | n/a |
What did safety testing find for Gemini 4 Carbon?
Safety level: Frontier Safety Framework (April 2026): full assessment is in the separate Gemini 4 Carbon Safety and Security Report; Google could not rule out that Argon and Carbon reached the biological CBRN CCL and the Cybersecurity CCL
- Versus Gemini 4 Argon, automated safety evaluations: text-to-text +3.50pp, multilingual +3.55pp, image-to-text -1.00pp, tone +0.10pp, unjustified refusals +0.10pp.
- Human red teaming: child safety evaluations met required launch thresholds, with similar or improved content safety versus Argon and no egregious concerns found.
- Google strengthened Frontier Safety safeguards for CBRN and cybersecurity misuse and provides specialised dual-use cyber capabilities only to verified defenders through the Fairwind Program.
- Gray Swan indirect prompt injection attack success is 0.0% versus 0.7% for Argon, per the launch post.
- Google says it is participating in the US government's voluntary pre-release model testing process.
Independent figures: Artificial Analysis, refreshed daily on our AI model benchmarks page. Full analysis: our Gemini 4 Carbon deep dive.
Google Model Timeline
1M tokens (up to 64K default synchronous output, up to 1M via background long decode) context
1M tokens (up to 64K default synchronous output, up to 1M via background long decode) context
1M tokens (combined input and output) context
1M tokens context
1,049k tokens context
66k tokens context
1,049k tokens context
33k tokens context
1,049k tokens context
1,049k tokens context
1,049k tokens context
8k tokens context
1,049k tokens context
1,049k tokens context
1,049k tokens context
8k tokens context
33k tokens context
1,049k tokens context
33k tokens context
96k tokens context
33k tokens context
131k tokens context
131k tokens context
96k tokens context
1,049k tokens context
1,049k tokens context
1,049k tokens context
8k tokens context
8k tokens context
Specifications
Related Tools
Google Veo
Google Veo is DeepMind's flagship text-to-video model, producing high-resolution, cinematic clips with strong prompt adherence and, in newer versions, synchronised audio. It is integrated across Google's creative and cloud products.
NotebookLM (Gemini Notebook)
Renamed Gemini Notebook in July 2026, this is Google's AI research and note-taking assistant that grounds its answers in your own uploaded sources. Known for source-grounded summaries, Audio Overviews, and, on paid tiers, a new sandboxed cloud computer that writes code and exports charts, slides and spreadsheets.
Claude Opus 5.5
Claude Opus 5.5, launched 22 September 2026, is Anthropic's current Opus-line flagship, replacing Opus 5. It performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, at $4 input / $20 output per million tokens (both 20% down, cache reads cut 60% to $0.20). It leads on Terminal-Bench 4.0, GDPval-AA v2.1 and Humanity's Last Exam, and is the first Opus model to ship with Fable-5.1-class cybersecurity, biology and distillation safeguards.
