Gemini 4 Carbon leads a broad AI index while access stays limited

Google announced Gemini 4 Carbon on 28 October, describing advances in cybersecurity and extended tasks. Its initial rollout is restricted to selected cyber defenders and trusted partners while safeguards are tested; the announcement does not establish general public availability. [1][4]

There is independent evidence for a broad performance gain. Artificial Analysis gave Carbon a score of 59 and first place in version 5 of its Intelligence Index. Its assessment also says Carbon does not lead in every area: it trails Claude Opus 5.5 in the separate Coding Agent Index and is not the leader in its knowledge-work evaluation. [2]

The apparent tension comes from measuring different things. The broad index combines several kinds of work, with coding, science, knowledge work and general tasks each receiving a quarter of its weight. Leading that combined score therefore does not mean winning every component or matching every user’s task. [2][3]

Google’s cybersecurity assertions need a further distinction. Its launch statement includes internal testing, whereas the outside results cited here concern defined benchmark suites. Those results cannot be treated as independent confirmation of every security claim or of safe deployment in a particular organisation. EWN has not tested the model. [1][2][3]

The evidence supports a narrower conclusion than an across-the-board victory: Carbon has a strong independently measured composite result and remains in a controlled rollout. Wider availability, task-specific performance and deployment safeguards remain separate questions. Google gives no firm public-release date in the announcement reviewed here. [1][2]