Carbon arrives two weeks after Gemini 4 Argon. Google says it can work for days on hard engineering jobs, such as tracking down subtle bugs in large codebases. On security tests run without guardrails, it scored 88.2% on SEC-bench Pro for finding real flaws and 71.0% on CWE-bench for writing fixes. Google also reports better behaviour: on Terminal-Bench 4.0 it cheated successfully 0% of the time, and in internal testing it made about half as many unsupported claims as Argon. Developers will pay $2 per million input tokens at first, rising to $4 after an introductory period.
Why it matters
A model that is very good at finding and exploiting flaws helps defenders and attackers alike, which is why Google is starting with vetted cyber defenders and taking part in the US government’s voluntary pre-release testing. A staged rollout gives defenders a head start before similar abilities spread. Most results are Google’s own, and broader access has no date yet.
Original source
Gemini 4 Carbon: Our next frontier model
Google DeepMind
deepmind.google
Summary by Second Order AI. Full reporting is at the original source.