Google's newest model, which it says is its strongest, will reach cyber defenders before anyone else. Gemini 4 Carbon was announced on 28 October 2026 in a blog post by Koray Kavukcuoglu, Google's Chief AI Architect and SVP of Google DeepMind, two weeks after Gemini 4 Argon. Google says Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59 and leads three cybersecurity tests: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploiting them and 71.0% on CWE-bench for patching them.
- Artificial Analysis Intelligence Index v5.059 Argon: 50, per Google's table
- SEC-bench Pro (finding vulnerabilities)88.2 % Google's figure
- ExploitGym (exploitation)54.8 % Computed by Google
- Introductory input price2 $ per million tokens $4 after the introductory period
- Introductory output price10 $ per million tokens $20 after the introductory period
| Benchmark | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| AA Intelligence Index v5.0 | 59 | 50 | 57 | 58 |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | n/a |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% |
| OSWorld-2.1 (offline subset) | 81.5% | 74.7% | 80.9% | 87.2% |
Source: Google DeepMind, Gemini 4 performance table, 28 October 2026. Google says non-Gemini scores are the providers' self-reported figures unless it states otherwise. n/a: no score shown.
Google's own comparison table also lists rivals' scores. On the Artificial Analysis index, Carbon's 59 compares with 58 for Anthropic's Claude Opus 5.5, 57 for OpenAI's GPT-6 Astra and 50 for Argon. Rivals come out ahead on some tests in the same table: GPT-6 Astra scores 65.5% and Claude Opus 5.5 scores 62.3% on FrontierSWE 2, a coding benchmark, against 61.2% for Carbon, and Claude Opus 5.5 scores 87.2% on an offline subset of OSWorld-2.1, a computer-use test, against 81.5%. Google's evaluation notes say it ran Carbon's ExploitGym test itself. Artificial Analysis lists the same score of 59 on its own page for the model.
A gradual rollout
Starting this week, Google says, Carbon goes to a subset of cyber defenders, with access widening gradually to more trusted partners through its Fairwind Program. The programme's page says it offers advanced models early to high-priority defenders, including governments, healthcare providers and telecoms services, and that it has more than 650 partners worldwide. Google also says it is taking part in the US government's voluntary process for testing models before release. It gave no date for wider access, saying it will keep gathering feedback and refining guardrails before offering Carbon to developers, businesses and consumers through its products.
Google says Carbon launches at Argon's introductory price: $2 for each million input tokens and $10 for each million output tokens, with a 95% discount on cached input. A footnote to the announcement says prices of $4 per million input tokens and $20 per million output tokens will apply once the introductory period ends, without saying when that is. Google also says Carbon more than doubled Argon's success rate at writing secure code in internal tests, and that its Chrome team found it more than 10 times better than another lab's large frontier model at producing high-quality patches. It adds that Carbon gained 89 Elo points to reach fourth place on WebDev Arena, with a score of 1768.
- Google blog: Gemini 4 Carbon, our next frontier model (Koray Kavukcuoglu), 28 Oct 2026, 18:00 CET
- Google DeepMind: Gemini 4 performance table
- Google DeepMind: Gemini 4 Carbon model evaluation (methodology)
- Google DeepMind: Fairwind Program
- Artificial Analysis: Gemini 4 Carbon model page