Google Unveils Gemini 4 Carbon, a Frontier AI Model the Public Cannot Yet Use

Google has announced Gemini 4 Carbon, a new frontier AI model, only two weeks after it rolled out Gemini 4 Argon. The company is planning a staged release for Carbon, and like the rollout of Argon, it could take a while before the public can use it.

The announcement follows a pause earlier this year in which Google did not release new frontier models. Carbon is not the Flash-class model that many have been waiting for. Instead, it is presented as an improvement on Argon.

As with earlier launches, Google says the model reaches a new threshold in coding, agentic tasks and scientific analysis. The company also says it is so capable that it cannot yet be made available to everyone.

Cybersecurity is where Google claims the biggest gains. The company describes a "significant leap" in the model's ability to detect, validate and fix vulnerabilities. It also says Carbon is twice as likely as Argon to produce code that is free of vulnerabilities.

The Chrome team adds that Carbon is 10 times better at generating "high-quality" patches than non-Google models. Google further says the model improved its score by 89 points on WebDev Arena, where it now ranks fourth on the leaderboard.

Google backs these broad assessments with benchmark figures that put Carbon at the top of the charts. It beats Argon by slim margins in every test the company has shared. Carbon also outperforms GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 in almost every test.

Some caution is warranted. Google presents Carbon as a completely new model, but it could be a post-training upgrade of Argon, given how much the two have in common. Benchmarks also do not tell the whole story, and every frontier model scores in roughly the same range on these tests. The gap between Carbon and its competitors amounts to just a few points.