Google announces Gemini 4 Carbon, built for multi-day agentic engineering runs

Google says Carbon will launch at the same introductory price as Argon, $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. For now it goes to "a subset of cyber defenders" and then to more trusted partners through the Fairwind Program, with developer and business access described as coming "as soon as possible." No date is given.

The post makes several claims that bear on unattended agent work. Google says its teams saw Carbon "run for days" on tasks such as root-causing and patching bugs in large codebases. It reports that successful cheating on Terminal-Bench 4.0 fell to 0.0%, with attempts at 0.4%, and that high-severity misalignment alerts in internal live coding environments dropped by almost 80%. It also reports 26% less overreach and 29% less under-service on multi-step plans, and about half as many unsupported claims as Argon. On security benchmarks it reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench, evaluated without guardrails. Google also cites Artificial Analysis Intelligence Index v5.0, where it says Carbon ranks first at 59. None of these figures has been independently checked, and the post does not publish the methodology behind the internal measurements.

Why it matters: The claims most relevant to a factory are the ones about shortcut-taking and staying in scope, because those are the failures a verification loop is built to catch. They are worth testing yourself once access opens. Until then there is nothing to adopt, and the pricing is only a plan.