Google Unveils Gemini 4 Carbon, Tops AI Intelligence Ranking

Google has unveiled Gemini 4 Carbon, a new AI model that tops the Artificial Analysis Intelligence Index with improvements in coding and cybersecurity. Access is initially limited to selected cybersecurity defenders.

Carbon scores 59 at its high reasoning setting on Artificial Analysis' v5.0 index, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, both at their maximum reasoning settings. The independently run index combines nine evaluations spanning coding, knowledge work, science, and general knowledge.

The model builds on Gemini 4 Argon, announced earlier this month. Google reports that internal teams have used Carbon for engineering tasks lasting days, including tracking down and patching bugs across large codebases.

In cybersecurity testing, it scored 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for validating them through exploitation, and 71.0% on CWE-bench for producing fixes. Those results were evaluated without guardrails.

Internal tests also showed roughly half as many unsupported claims as Argon and nearly 80% fewer high-severity misalignment alerts in live coding environments. Carbon resisted every attack in the evaluated Gray Swan indirect prompt-injection benchmark, though that does not establish immunity to attacks outside the test.

Access will expand through the Fairwind Program before reaching developers, businesses, and consumers. No general release date has been announced.

Introductory pricing will be $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached input. After that period, input and output prices rise to $4 and $20, respectively.