Google announces Gemini 4 Carbon, its new frontier AI model

Google's claim of Gemini 4 Carbon's leadership is supported by its reported performance across various benchmarks. According to Google's own comparisons, Carbon outperforms its predecessor, Argon, along with OpenAI's GPT-6 Astra, and Anthropic's Claude Opus 5.5 and Fable 5.1 on 15 of 17 internal tests. These results suggest an advancement in areas such as web development, where Carbon has climbed to fourth place on the WebDev Arena leaderboard.

A notable shift in Google's reporting is the use of Artificial Analysis, an independent testing firm, for Terminal-Bench 4.0 scores. This benchmark measures an AI model's ability to perform real technical tasks by independently issuing commands in a computer terminal. On this challenging test, Carbon reportedly improved by approximately eight points over Argon and surpassed Claude Opus 5.5, addressing previous inconsistencies between company-reported and independent scores.

Despite these reported gains, the practical availability of Gemini 4 Carbon remains limited. Google has announced the model well before a public release, with initial access restricted to a small group of cybersecurity defenders. The Artificial Analysis Intelligence Index, where Carbon currently leads by a narrow one-point margin over Claude Opus 5.5, also flags Carbon as lacking a publicly available API, meaning it cannot yet be used outside Google’s internal access programs.