Google announced Gemini 4 Carbon on October 28, 2026, as a model for cybersecurity and sustained engineering work. Its announced access plan starts with a subset of cyber defenders through the Fairwind Program, followed by trusted partners and, later, developers, businesses and consumers.
Who Google planned to give access first
The announcement described a staged rollout, not general availability. Google said it planned to begin with selected cyber defenders, then extend access to trusted partners before broadening it to developers, businesses and consumers.
What the cybersecurity scores measure
Google reported three results for different security tasks. It said the external evaluations were conducted without guardrails.
| Benchmark | Score reported by Google | Task evaluated |
| SEC-bench Pro | 88.2% | Vulnerability discovery |
| ExploitGym | 54.8% | Exploitation evaluation |
| CWE-bench | 71.0% | Vulnerability patching |
These percentages describe separate evaluations, rather than one overall cybersecurity score. Google also reported a score of 59 and first place for Carbon on Artificial Analysis Intelligence Index v5.0.
That index result does not mean Carbon led every task. VentureBeat reported that GPT-6 Astra scored higher than Carbon on FrontierSWE 2, while Claude Opus 5.5 scored higher on the OSWorld offline subset. The reported results put Carbon at 61.2% on FrontierSWE 2, compared with 65.5% for Astra, and at 81.5% on the OSWorld subset, compared with 87.2% for Opus.
What the PDF guidance says about Carbon
Google's general Gemini PDF guide describes processing documents up to 1,000 pages, while its Gemini 4 model page lists PDF input for Gemini 4 Argon. Those are general Gemini and Argon-specific details, respectively, not Carbon specifications. For background on Argon, see NeoTeo's earlier coverage of Gemini 4 Argon and practical coding questions.
