Google's Gemini 4 Carbon goes first to cyber defenders
Google announced Gemini 4 Carbon, an AI model it says tops a leading intelligence index and excels at finding and fixing security flaws. It goes first to cyber defenders and trusted partners, with no public release date.
AI-assisted · US Brief desk · Sources listed below
What happened
Google announced Gemini 4 Carbon on Oct. 28, two weeks after releasing Gemini 4 Argon, in a blog post from Google DeepMind's Koray Kavukcuoglu. Google said the model ranks No. 1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59; The Verge reported that puts it ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra.
The headline claim is security. Google reported scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, measured without safety guardrails, and said Carbon writes more than twice as much secure code as Argon. Google said its Chrome team found Carbon produced more than 10 times better patches than "another lab's large frontier model," which it did not name. It also said the model climbed 89 Elo points to No. 4 (1,768) on WebDev Arena.
Key numbers: On safety, Google said Carbon cheated 0.0% of the time on Terminal-Bench 4.0, with cheating attempts at 0.4%; high-severity misalignment alerts fell about 80%; unsupported claims were about half as frequent; overreach fell 26% and under-service 29%; and its rate on Gray Swan's indirect prompt injection test was 0.0%, versus 0.7% for Argon. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off, rising to $4 and $20 after the introductory period.
Why it matters
A model good enough at finding and exploiting software flaws can help defenders patch faster, but the same skill could help attackers. Google's decision to restrict the first release is the latest sign that leading AI labs are treating cyber capability as a reason to stage launches rather than open them to everyone at once.
What’s next
Independent tests from outside researchers once access widens, and whether Google sets a date for developers.
More context
Rollout: Google said a subset of cyber defenders gets access this week, followed by trusted partners through its Fairwind Program, and that it took part in US government voluntary pre-release testing. It gave no date for availability to developers or consumers.
What's still unknown: None of the security or safety figures has been independently reproduced yet, and Google did not name the rival model it compared against in the Chrome test. It also has not said how it will decide when the model is safe for wider release.
3 listed sources
References listed by US Brief; a source count is not a verification score.
Editorial sourcing notes
Sourcing: Google's announcement blog post from Koray Kavukcuoglu, The Verge and 9to5Google, Oct. 28, 2026. All benchmark, safety and pricing figures are Google's own claims unless noted; the Artificial Analysis ranking is as reported by Google and The Verge. Photos are file photos of Google's Mountain View campus from 2009 (hero) and 2014 (body).
Like US Brief? You can support it with a tip.