AI Race: Google Releases Gemini 4 Carbon Without Public Access
The new model outperforms rivals in coding and security tasks, but Google is holding back general availability for a staggered rollout.
By Marco Graham ·
Google announced Gemini 4 Carbon on Tuesday, a frontier AI model that leads industry benchmarks but remains unavailable to most users. The release follows Gemini 4 Argon by just two weeks, signaling a rapid pace of development after a pause earlier in the year.
The company states that Carbon achieves major gains in cybersecurity. Google reports a significant leap in vulnerability detection, validation, and remediation capabilities. The model is twice as likely as Argon to produce code free of security flaws.
Performance in web development has also improved. The Chrome team claims Carbon is ten times better than non-Google models at generating high-quality patches. The model also gained 89 points on the WebDev Arena leaderboard, moving it to the number four position.
Google positions Carbon as a distinct new model, though it shares many characteristics with Argon. In internal tests, Carbon beats Argon by small margins across all categories. It also outperforms GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 in almost every metric provided.
Despite these rankings, the gap between Carbon and its closest competitors is narrow, with all frontier models scoring in the same range. Google is implementing a staged release strategy for Carbon, a process that could take time before the model becomes widely accessible.
