Google opens Gemini 4 Carbon to selected cyber defenders

Google announced Gemini 4 Carbon on 28 October and is starting access with selected cyber defenders, followed by trusted partners in its Fairwind programme. Wider developer, business and consumer availability has no announced date.[1]

Artificial Analysis lists Carbon’s high-reasoning variant at 59 on its Intelligence Index v5.0, first among 52 models in its summary. The index combines nine evaluations spanning knowledge work, coding, science and reasoning; the lead is a composite result.[3]

The cybersecurity headline covers different jobs. Google reports 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation and 71% on CWE-bench for patching, with guardrails off for these tests.[1]

What counts as success

For SEC-bench Pro, Google tested 183 tasks from the May 2026 JavaScript-engine split and averaged three independent attempts per task. A result could qualify by crashing the vulnerable build but not the patched one; when both crashed, a model judge checked whether the intended vulnerability caused it.[2]

ExploitGym used Google’s Antigravity agent setup without internet access. Success required capturing a target flag and a model judge verifying exploitation of the specified flaw. The methodology reports both six-hour and ten-hour timeouts. CWE-bench scores came from Collinear AI’s own evaluations. These are separately defined tests, not three measurements of one skill.[2]

Two prices, two stages

Google advertises introductory rates of $2 per million input tokens and $10 per million output tokens, then $4 and $20 after the introductory period. It gives no end date for that period.[1]

Artificial Analysis displays the later $4/$20 rates. Its model page also labels the API as not publicly available. A listing with performance and pricing should therefore be read alongside the access restriction: an evaluated model can appear in a comparison before a general customer can use it.[3]

Report an error or suggest a correction