AIO APEX
en

Google’s Gemini 4 Carbon tops AI benchmarks with cybersecurity as its core capability

Google Blog
Share:
Google’s Gemini 4 Carbon tops AI benchmarks with cybersecurity as its core capability

Google announced Gemini 4 Carbon on October 28, calling it “our strongest model with state-of-the-art cybersecurity and long horizon capabilities.” The model scores 59 on the Artificial Analysis Intelligence Index v5.0, placing it first, and represents a significant departure from previous Gemini releases in that its primary design focus is defensive security rather than general capabilities.

The numbers Google published on cybersecurity benchmarks are striking. On SEC-bench Pro, which tests vulnerability finding, Carbon scores 88.2%. On ExploitGym, which measures exploitation ability, it scores 54.8%. On CWE-bench, which evaluates patch quality, it reaches 71.0%. Internally, Google says Carbon generates secure code at more than twice the success rate of its predecessor Argon. The Chrome team tested Carbon against another unnamed large frontier model and reported over 10x better high-quality patches.

Carbon also marks progress on alignment and safety. Cheating behavior on Terminal-Bench 4.0 dropped to 0.0%, and cheating attempts to 0.4%. High-severity misalignment alerts in live coding sessions fell by nearly 80% compared to Argon. Unsupported claims were cut roughly in half, overreach fell 26%, and under-service dropped 29%. Its Gray Swan prompt injection success rate is 0.0%, compared to 0.7% for Argon.

The model is also built for sustained multi-step agentic work. Google says internal teams have run Carbon for days on complex engineering problems without interruption — a requirement that matters for security workflows that involve chained analysis across large codebases or infrastructure.

Access is deliberately narrow at launch. Carbon is currently available to a selected subset of cybersecurity defenders, with broader access rolling out through Google’s Fairwind Program before it reaches developers, businesses, and consumers. Pricing is set at $2 per million input tokens and $10 per million output tokens at the introductory rate, rising to $4 and $20 per million tokens respectively afterward, with cached input priced at a 95% discount.

The restricted launch is significant. Google says it limited early access because of Carbon’s advanced cybersecurity capabilities — an acknowledgment that a model this capable of finding and exploiting vulnerabilities requires careful controls around who can use it and for what purpose.

Carbon arrives two weeks after Google launched Gemini 4 Argon, suggesting Google is moving on an accelerated release cadence for specialized variants. As reported by Google’s official blog, authored by Koray Kavukcuoglu.

Originally reported by Google Blog. Read the original article for additional details.

View original source
Share: