Google DeepMind announced Gemini 4 Carbon on 28 October, describing it as its strongest model for cybersecurity and long-horizon work. It ranks first on v5.0 of the Artificial Analysis Intelligence Index at 59. On external cyber benchmarks, evaluated without guardrails, Google reports 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitation and 71.0% on CWE-bench for patching. Internal comparisons give more than 2x the secure-code success rate of Gemini 4 Argon, and the Chrome team reports Carbon producing high-quality patches more than 10x better than "another lab's large frontier model". Carbon lands at fourth on WebDev Arena at 1768, up 89 Elo. Rollout is staged: from this week a subset of cyber defenders get access through the Fairwind Program, with developers, businesses and consumers to follow; Google says it is taking part in the US government's voluntary pre-release testing process. List price is unchanged from Argon at $2 per million input tokens and $10 per million output, with cached input at 95% off.
Our takeAll three headline numbers come from benchmarks Google selected, and the post says they were run without guardrails, so they measure capability rather than deployed behaviour. The 2x and 10x comparisons are internal, with no test set released, and staged access means outside researchers cannot reproduce the scores yet. The long-horizon claim (models running for days) is the hardest to check: multi-day agent runs need a stated success criterion and variance, not a single internal anecdote.