On cybersecurity benchmarks, Google reports Carbon leading SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% for turning discovered flaws into working exploits, and CWE-bench at 71.0% for writing patches. Internally, the company says Carbon generates secure code more than twice as often as Argon, and its Chrome team reports it produced high-quality patches more than ten times better than a rival lab's frontier model. Those internal figures come from Google itself and cannot be independently verified.

Carbon is built for sustained agentic workflows: Google says its teams have seen the model run for days, working through dead ends on tasks such as root-causing and patching subtle bugs across large codebases. It also jumps 89 Elo points to fourth place on WebDev Arena, and Google reports gains in scientific reasoning, multimodal understanding and full-stack web development.

Independent evaluations add useful nuance. The Decoder notes that on Vals AI Carbon places second at 70.36% accuracy, behind Claude Sonnet 5.5, and that in agentic coding it still trails Claude Code with Opus 5.5. Its answer accuracy rises from 48% to 58% compared with Argon, but the hallucination rate climbs from 4% to 11%, and the Finance and Tax Agent benchmark takes more than twice as long as with Argon, while Legal Research tasks average more than 50 minutes each.

Access is deliberately staged. Starting this week, a subset of cyber defenders gets access, with wider availability to trusted partners through Google's Fairwind Program, while the company participates in the U.S. government's voluntary pre-release testing process. Pricing matches Argon at $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% off.