Gemini 4 Carbon leads on security benchmarks and long agentic runs

Google released Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon, and the pitch is narrow on purpose: security work and long agentic runs. Koray Kavukcuoglu, who leads Google DeepMind, says internal teams have watched it run for days on a single task, working through dead ends to root-cause and patch bugs in large codebases. The benchmark numbers back the security claim. Carbon takes the top spot on the Artificial Analysis Intelligence Index v5 with a score of 59, and on security evals run without guardrails it reaches 88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym for exploitation, both described as state of the art. Google's Chrome team reported more than 10x better high-quality patches than a rival lab's frontier model.

The alignment numbers are the other half of the story, and they come mostly from internal testing. Google says successful cheating on Terminal-Bench 4.0 fell to zero, high-severity misalignment alerts in live coding environments dropped by almost 80%, and prompt injection attack success on the Gray Swan benchmark was 0.0% against 0.7% for Argon. Pricing starts at $2 per million input tokens and $10 output during an introductory period, rising to $4 and $20 after, with cached input 95% cheaper. The catch is access: it goes first to a subset of cyber defenders and expands only gradually through Google's Fairwind program, with no firm date for general availability.

Why it matters

If you work in security, Carbon is being handed to defenders first and tuned for exactly the vulnerability-finding and patching you do, so it is worth early testing. For everyone else, the gated rollout means you cannot build on it yet, and the headline alignment numbers come from Google's own tests, not independent ones.