Google’s Gemini 4 Carbon tops the intelligence rankings and finds, exploits, and patches its own vulnerabilities

The cybersecurity benchmarks are where the model is clearly built to live. Tested without guardrails, Carbon hit 88.2% on SEC-bench Pro for vulnerability discovery, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. Google says it generates secure code at more than twice Argon’s success rate internally, and Chrome’s own team reported Carbon produced high-quality patches over ten times more often than a comparable frontier model from another lab. Security firm Wiz reports it leads on web penetration testing benchmarks too. That combination of skills, finding holes, exploiting them, and fixing them, all in the same model, sits at an odd intersection for a tool meant to defend networks rather than break into them, not unlike the dual-use questions that came up around OpenAI’s own vulnerability-scanning tools.

Google frames the long-horizon strength the same way: internally, Carbon has run for days at a stretch on multi-step engineering tasks, working through dead ends while root-causing and patching subtle bugs buried in large, low-level codebases, the kind of grinding work that used to need a human engineer’s sustained attention across a workweek, not a single prompt.

Pricing during the introductory period runs $2 per million input tokens and $10 per million output tokens, cached input discounted 95%, before both rates roughly double once that window closes. Access right now is limited to a subset of cyber defenders, with Google planning to widen it gradually to trusted partners through its Fairwind Program before any broader developer or consumer release, no firm date given.

On the safety side, Google reports several numbers moving in the right direction compared to Argon: cheating attempts on Terminal-Bench 4.0 dropped to 0.4% with zero successful attempts, high-severity misalignment alerts in internal live coding fell by nearly 80%, and the model’s attack success rate on Gray Swan’s indirect prompt injection benchmark came in at 0.0%, down from Argon’s 0.7%. Google also says it’s participating in the US government’s voluntary pre-release testing process for frontier models, the same review track its recent Argon and Genie 4 releases went through.