On October 28, Google announced Gemini 4 Carbon, its most capable model to date — and arguably the most focused AI system ever aimed at cybersecurity. Built on the Gemini 4 Argon foundation, Carbon is purpose-engineered for multi-day security and engineering tasks that current models choke on. If you work in infosec or enterprise software, this one deserves your attention.
What Carbon Is Built For
Where most frontier models are generalists, Gemini 4 Carbon's entire training emphasis is on finding, exploiting, and patching software vulnerabilities. Google designed it to sustain multi-step reasoning over long sessions — the kind of tedious, iterative work that breaks less capable models after a few turns. The model is already in limited deployment through Google's Fairwind Program, where selected cyber defenders are getting early access before a broader rollout.
The Benchmarks
Google's internal numbers are striking. On SEC-bench Pro — a standardized benchmark for software flaw detection — Carbon scored 88.2%, substantially ahead of any previously published model. On ExploitGym (exploitation tasks) it hit 54.8%, and on CWE-bench Pro (patch quality) it reached 71.0%.
On the broader Artificial Analysis Intelligence Index v5.0, Carbon scored 59 — placing it first, ahead of Claude Opus 5.5 and GPT-6 Astra. That's the general intelligence leaderboard, not just the security-specific one.
Google's Chrome security team also reported that Carbon produced more than ten times higher-quality patches than an unnamed competing model. That comparison can't be verified independently since Google didn't name the rival.
Safety and Alignment
Cybersecurity capability and alignment are deeply intertwined — a model that can find vulnerabilities can also explain how to exploit them. Google says Carbon shows an 80% reduction in high-severity misalignment alerts compared to Argon, suggesting the additional training tightened rather than loosened its safety margins. The US government voluntary pre-release testing program evaluated the model before launch, per Interesting Engineering.
Access and Pricing
Right now, access to Carbon is restricted to selected defenders through the Fairwind Program — wider availability is planned but unscheduled. API pricing is set at $2 per million input tokens and $10 per million output tokens — expensive for a development tool, competitive for enterprise security tooling.
What This Means in Practice
The gap between general-purpose LLMs and security-specific tooling has been shrinking fast. Carbon's numbers suggest it's crossed a threshold where AI-assisted code audit is no longer just a novelty — it might be faster and cheaper than a full manual security review for routine work. The limiting factor is now access, not capability. Once Fairwind opens up, expect every major security vendor to integrate or compete with it.
Whether the benchmark claims hold up outside Google's own evaluations is still an open question. Independent red teams haven't published results yet. But the trajectory is clear: the best AI system for finding vulnerabilities in your software is now, apparently, an AI system built specifically for that job.