Google announced Gemini 4 Carbon on 28 October to provide significant gains in coding, agentic tasks, and cybersecurity. The model ranks number 1 on the Artificial Analysis Intelligence Index v5.0 and holds the top spot on SEC-bench Pro.
Carbon scored 59 on the Artificial Analysis Intelligence Index v5.0, 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71% on CWE-bench. Google stated that Gemini 4 Carbon shows reduced misalignment and fabrication rates in live coding environments while cutting unsupported claims by approximately half compared to Argon.
The model reduced under-service by 29% during multi-step workflows and shows an 89-point jump to fourth place on the WebDev Arena leaderboard. Google claims the model is built for completing longer tasks, avoiding shortcuts, resisting prompt injection attacks, and collaborating with users.
Introductory pricing for the model is $2 per 1 million input tokens and $10 per million output tokens, while standard pricing is $4 for input and $20 for output. Cached input tokens receive a 95% discount.
Gemini 4 Carbon is currently only available to a select group of cybersecurity defense companies this week before expanding to other partners in its Fairwind Program.
