Google introduces Gemini 4 Carbon model with a focus on cybersecurity

Summary

Google released Gemini 4 Carbon, the cybersecurity-focused successor to Argon, on October 29, 2026. The company claims the model tops the Artificial Analysis Intelligence Index v5.0 and outperforms GPT-6 Astra and Claude Opus 5.5 in vulnerability detection and exploit generation benchmarks. Currently, the model is available only to a limited number of defense sector partners under the Fairwind program.

Why it matters

Positioning a large language model (LLM) explicitly around offensive and defensive cyber capabilities signals a shift from general-purpose chatbots to specialized security tools. This has tangible implications for defense and enterprise IT procurement. By restricting access to a limited group of partners rather than the general public, Google demonstrates that it treats near-offensive capabilities as a controlled, trust-based service rather than a standard API. The pricing, set at $2/$10 per million tokens similar to Argon, keeps the barrier low for qualified buyers. If the Fairwind deployment expands, this capability could become accessible to mid-sized security teams as well.

Highlights

• It achieved top rankings in SEC-bench Pro and ExploitGym benchmarks, which measure real-world vulnerability detection and conversion to exploits.
• Internal multi-day engineering tests indicate that the model continues to make progress on complex tasks rather than giving up when it hits dead ends.
• It halved unsupported claims compared to Argon and added resistance to prompt injection, addressing two key points of corporate trust deficits.

Critical take

All performance claims in this text rely solely on Google's own press release. The cited benchmarks (SEC-bench Pro, ExploitGym) have not been independently audited or reproduced by third parties. Consequently, the framing of 'outperforming OpenAI and Anthropic' remains an unverified marketing statement. Because the model is accessible only to a handful of defense partners, no external security researchers or competitors have yet been able to stress-test its offensive capabilities to the limit. This leaves the real-world gap between benchmark scores and operational security value entirely unclear.