Google DeepMind unveils Gemini 4 Carbon frontier model

Google DeepMind introduced Gemini 4 Carbon on 28 October 2026, two weeks after releasing Gemini 4 Argon. Carbon is trained to handle multi-step software engineering, defensive cybersecurity, and long-horizon agent workflows while keeping Argon’s inference footprint and token pricing. On the independent Artificial Analysis Intelligence Index v5.0, Carbon ranks first overall with a score of 59, alongside an 89-point Elo gain over Argon on WebDev Arena to reach fourth place at 1,768.

Cybersecurity and defensive code generation

To test dual-use security capabilities before broad release, Google DeepMind evaluated an unguarded research checkpoint across three offensive and defensive benchmarks: SEC-bench Pro for vulnerability detection, ExploitGym for end-to-end exploit construction, and Collinear’s CWE-bench for root-cause patching. Carbon also more than doubled Argon’s score on Google DeepMind’s internal secure code generation suite, which tests whether a model introduces exploitable flaws when writing production patches.

Rollout and pricing

Because of its exploit-generation scores, Google DeepMind is staging deployment through the Fairwind Program, giving vetted security teams, critical-infrastructure maintainers, and trusted enterprise partners early access for defensive red-teaming and automated vulnerability remediation before general availability on Vertex AI and Google AI Studio.

Key numbers

• 59: Rank #1 score on the Artificial Analysis Intelligence Index v5.0.
• 88.2% / 54.8% / 71.0%: Scores on SEC-bench Pro, ExploitGym, and CWE-bench when tested without guardrails.
• $2 / $10 per million tokens: Introductory input and output API pricing, matching Gemini 4 Argon, with a 95% discount on cached input tokens.