Google releases Gemini 4 Carbon to cyber defenders first and says it leads key benchmarks for finding software flaws

Google has released Gemini 4 Carbon, which it describes as its strongest model, but for now only to a selected group of cyber defence organisations. In an announcement on 28 October, Koray Kavukcuoglu of Google DeepMind said access would widen in stages to trusted partners in the company's Fairwind Program, and later to developers, businesses and consumers once Google judges it safe. No date was given. Google also says it is taking part in the US government's voluntary pre-release testing of models.

By Google's account, Carbon scores 59 and ranks first on the Artificial Analysis Intelligence Index v5.0; The Verge reports that Claude Opus 5.5 and GPT-6 Astra follow it. On cyber tests run without guardrails, Google reports 88.2% on SEC-bench Pro, which measures finding real vulnerabilities, 54.8% on ExploitGym, which measures turning flaws into working exploits, and 71.0% on CWE-bench for patching. It says internal tests show more than double the success rate of Gemini 4 Argon, released two weeks earlier, at writing secure code.

Google says the model can keep working on multi-step engineering tasks for days. On safety, it reports almost 80% fewer high-severity misalignment alerts in internal use, about half as many unsupported claims as Argon, and no successful attacks on Gray Swan's indirect prompt injection benchmark. All of these are the company's own measurements.

Introductory pricing matches Argon at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period.