Google has introduced Gemini 4 Carbon, its second frontier AI model in two weeks following Gemini 4 Argon. Carbon is not the smaller Flash-style model many users have been waiting for; instead, Google describes it as an improved version of Argon. As with Argon, the release is being rolled out gradually and may take some time to become broadly available.
Google says its newest models are setting new benchmarks in coding, agent-based tasks, and scientific analysis. However, the company has not yet made all of Carbon’s capabilities available for general use.
Claims of stronger cybersecurity performance
Google says Gemini 4 Carbon delivers a “significant leap forward” in vulnerability detection, verification, and remediation. The company also says Carbon is twice as likely as Argon to generate code that does not contain vulnerabilities.
The Chrome team claims that Carbon produces “high quality” patches 10 times more often than models from outside Google. Google also says Carbon’s score increased by 89 points. On the WebDev Arena leaderboard, Carbon is currently ranked fourth.
Benchmark results
Google has published benchmark figures to support these claims, with Carbon at the top of the chart. The company presents Carbon as an entirely new model, although it is also possible that the system is a post-training upgrade to Argon.
The two models have many similarities, but Carbon narrowly surpassed Argon in all of the tests conducted for the comparison. The new model also outperformed GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1 in nearly every test.
Benchmarks do not capture every aspect of a model’s performance, and the results show relatively small differences among the leading systems. All of the frontier models achieved roughly similar scores, with only a few points separating Carbon from its competitors.
Source: arstechnica.com
