Google Built an AI Hacker. Why Is Access So Locked Down?

Just two weeks after shipping the Argon model, Google DeepMind launched Gemini 4 Carbon, a new frontier model that claims the top spot on the Artificial Analysis Intelligence Index v5 with 59 points, according to reports from The Verge and The Decoder. Unlike typical general-purpose releases, Carbon is explicitly designed for extended, multi-day engineering workflows and the full cyber defense lifecycle.

What Makes Carbon Different in Practice?

According to evaluations cited by The Decoder, Carbon achieved an 88.2 percent score on SEC-bench Pro, a benchmark measuring autonomous vulnerability discovery in real-world software, and 54.8 percent on ExploitGym for turning bugs into exploits. Google ran these specific security tests without safety guardrails to evaluate raw capabilities.

Internal teams reported running the model for days without human intervention to push through dead ends in complex, multi-step codebases. Furthermore, accuracy on reasoning tasks rose to 58 percent compared to Argon's 48 percent, while maintaining competitive pricing at $2 per million input tokens and $10 per million output tokens.

How to Access Gemini 4 Carbon

Because of its powerful capabilities in finding vulnerabilities and generating patches, Google is enforcing strict access controls rather than opening it immediately to the public.

• Initial rollout: Available exclusively to a select group of cybersecurity defense companies and experts starting this week.
• Government pre-testing: The model is undergoing voluntary pre-release testing with U.S. regulatory frameworks.
• Fairwind Program: Access will gradually expand to trusted enterprise partners through Google's Fairwind Program before any consumer interfaces launch.

Is It Worth It for Development Teams?

Carbon offers a strong alternative to Anthropic's Claude Opus 5.5, delivering high analytical and presentation quality at a lower task cost (roughly $7.76 per task). However, its heavy resource usage — averaging over 50 minutes per task in legal research and generating 2.3 times more reasoning tokens than Argon — means it is optimized strictly for deep, heavy engineering tasks rather than quick conversational queries.