Google launches Gemini 4 Carbon, its latest frontier AI model

Google announced Gemini 4 Carbon, its newest frontier AI model, in a blog post on Oct. 28. The model, described by Koray Kavukcuoglu, senior vice president of Google DeepMind, is aimed at cybersecurity applications and complex, multi-step tasks.

Gemini 4 Carbon currently holds the top position on the Artificial Analysis Intelligence Index, edging out Anthropic’s Claude Opus 5.5 by one point. In independent testing by Artificial Analysis, the model achieved the highest scores on 15 of 17 benchmark tests, surpassing competitors such as GPT-6 Astra and previous Google models.

The model is being rolled out initially to a limited group of cybersecurity defenders through Google’s Fairwind Program, with broader access to developers and consumers planned for a later date. Pricing for the API, when released, is expected to match the introductory rates for the earlier Argon model, at $2 per million input tokens and $10 per million output tokens.

"Carbon uses the same safety system as Argon and adds additional alignment work, resulting in fewer misalignment alerts and unsupported claims," Google said in its testing report.

Google noted that Carbon did not succeed in cheating on the Terminal-Bench 4.0 test and blocked all prompt-injection attempts, indicating improvements in safety and reliability over its predecessor.