Google says Gemini 4 Carbon excels at finding software flaws

Google has unveiled Gemini 4 Carbon, a frontier AI model from its Gemini line that Google says leads the field in identifying, exploiting and patching software vulnerabilities. Carbon scored 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym’s exploitation benchmark, and 71.0% on Collinear’s CWE-bench patching benchmark, outperforming rival models including OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 and Claude Fable 5.1 in the company’s internal comparisons.

Google notes these results were achieved with no guardrails in place. The model is not yet broadly available; access is being rolled out first to vetted cyber defenders via the Fairwind Programme, with plans to widen access to approved partners. Google says more than 650 organisations participate in Fairwind, which grants early access to advanced models for cyber defence to governments, critical infrastructure operators and major technology platforms.

Carbon’s performance is framed against its predecessor, Gemini 4 Argon, with some Fairwind partners already having exclusive access to Argon’s cybersecurity capabilities. Google says Carbon is more than twice as successful as Argon at writing secure code and, in generating high-quality patches, outperformed a large frontier model from a rival lab by more than tenfold, according to Google’s internal benchmarks.

Wiz, Google’s cloud security firm, has observed Carbon achieving top results on web penetration testing benchmarks. In terms of safety, Carbon showed resilience in Gray Swan’s indirect prompt injection benchmark (no successful attacks) compared with Argon’s 0.7% success rate; on Terminal-Bench 4.0, Carbon’s cheating attempts also failed.

Google notes ongoing collaboration with the US government on voluntary model vetting before wider release, with Carbon priced at introductory rates of $2 per million input tokens and $10 per million output tokens, moving to $4 and $20 later.