Google DeepMind Introduces Gemini 4 Carbon, Its Top Model for Cyber Defense and Coding

Google DeepMind has introduced Gemini 4 Carbon, a frontier model built for long-running coding work and cybersecurity defense. The company says it is the strongest model it has released so far. It also tops the updated v5.0 Artificial Analysis Intelligence Index with a score of 59.

Carbon is the second Gemini 4 model after Gemini 4 Argon, and Google is not opening it up broadly. Access starts with a vetted group of cyber defenders, then moves through the Fairwind Program to governments, critical infrastructure operators, and core tech platforms. Everyone else has to wait, and Google did not give a date for wider release.

On the benchmark side, Carbon does look sharp. Google says it leads 15 of 17 published benchmarks against Argon, GPT-6 Astra, Claude Opus 5.5, and Fable 5.1. The biggest standout is GraphWalks, where Carbon scores 96.2% across the 256K to 1M token range, well ahead of Argon’s 84.2%. It also posts 65.2% on Terminal-Bench 4.0, 61.2% on FrontierSWE, and 77.4% on Agents’ Last Exam v1.1.

For cybersecurity, Google is pitching Carbon as a model that can find flaws, prove them, and patch them. It gets 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on Collinear’s CWE-bench. Google says those tests were run without guardrails, and that the model showed more than 2x higher secure-code success than Argon in internal work.

The safety story is stronger than the usual launch gloss. Google says Carbon keeps Argon’s layered defenses and that several misalignment measures improved, including a drop to 0.0% on successful cheating in Terminal-Bench 4.0. Still, most people can’t touch it yet. The pricing is public, the model is not: $2 per million input tokens and $10 per million output tokens at intro rates, rising later to $4 and $20.