Google unveils Gemini 4 Carbon, a frontier model aimed at cyber defence — first access goes only to vetted defenders

Google DeepMind announced Gemini 4 Carbon on 28 October 2026, two weeks after Gemini 4 Argon. The company calls it its strongest model, focused on coding, long-running agentic work and cybersecurity defence.

What Google claims

• #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59, ahead of Claude Opus 5.5 and GPT-6 Astra, according to Google and The Verge.
• SEC-bench Pro 88.2% (finding real vulnerabilities), ExploitGym 54.8% (checking whether a flaw is exploitable) and Collinear’s CWE-bench 71.0% (writing working fixes). Google notes these scores were measured without guardrails.
• More than 2x the success rate of Argon at generating secure code on internal tests; the Chrome team reported far better patches than a rival lab’s model.
• On Gray Swan’s indirect prompt-injection benchmark, no attack succeeded against Carbon (vs 0.7% for Argon), SecurityWeek reported.

Staged rollout

Because the same skills could help attackers, Google is starting with a subset of cyber defenders this week and will widen access through its Fairwind Program (over 650 organisations, per SecurityWeek). It also takes part in the US government’s voluntary pre-release testing. Broader developer and consumer access will follow later. Introductory price: $2 per million input tokens and $10 per million output tokens.

Caveat

All benchmark figures come from Google and have not yet been independently reproduced.