Google has introduced Gemini 4 Carbon, a new frontier AI model released only two weeks after Gemini 4 Argon. Carbon is not the long-awaited Flash equivalent; instead, Google describes it as an improved version of Argon. The company is rolling it out gradually, and access may remain limited while Google evaluates its performance in coding, agent tasks, scientific analysis, and cybersecurity.
Google highlights cybersecurity and coding gains
Google says Gemini 4 Carbon marks a “significant leap forward” in vulnerability detection, verification, and remediation. According to the company, Carbon is twice as likely as Argon to generate non-vulnerable code.
The Chrome team also claims that Carbon is 10 times better than non-Google models at producing “high quality” patches. Google says Carbon’s score increased by 89 points, although WebDev Arena currently places it fourth on its leaderboard.
Benchmark results remain close
Google has published figures to support its performance claims, with Carbon at the top of the company’s comparison chart. However, Google is presenting Carbon as an entirely new model, and it remains unclear whether it could instead be a post-training upgrade to Argon.
The two models share many similarities, but Carbon narrowly outperformed Argon in all of the reported tests. The new model also surpassed GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 in nearly every benchmark.
Those results should be treated cautiously because benchmarks do not capture every aspect of an AI model’s real-world performance. The frontier models tested by Google achieved broadly similar scores, with only a few points separating Carbon from its competitors. Despite the company’s claims about its capabilities, Google has not yet made the model broadly available.
Source: arstechnica.com