Google unveils Gemini 4 Carbon AI model to detect code vulnerabilities

According to Google, Carbon scored 59 points in the independent Artificial Analysis Intelligence Index v5.0 ranking, taking the top spot, followed by Claude Opus 5.5 with 58 points. In internal company testing, the model ran for several consecutive days without human intervention: analyzing large codebases, tracking down root causes of bugs, preparing patches, and pivoting after unsuccessful attempts.

In security benchmarks, Carbon scored 88.2% on vulnerability detection in SEC-bench Pro (compared with 83.4% for the predecessor Gemini 4 Argon), 54.8% on exploitability assessment in ExploitGym (42% for Argon), and 71% on repairing vulnerable code in CWE-bench (68% for Argon). Google notes that these measurements were conducted without the model's safety guardrails. In the WebDev Arena ranking, the model rose to fourth place with 1,769 points (1,768 according to Kod.ru). However, leadership varies across tests: in FrontierSWE 2, the new Google model is outperformed by GPT-6 Astra and Claude Opus 5.5.

Carbon will initially be made available to a limited circle of cyber defense specialists, after which Google will expand access through the Fairwind program. The model will open to developers, businesses, and regular users following additional verification of its safety mechanisms; a general launch date has not been specified. The introductory API price is $2 per million input tokens and $10 per million output tokens, with cached input tokens being 95% cheaper. After the introductory period, whose duration the company has not specified, pricing will rise to $4 and $20, respectively.