Google's Gemini 4 Carbon is its strongest model yet — and cyber defenders get it first

Google announced Gemini 4 Carbon on October 28, just two weeks after Gemini 4 Argon. It's the company's most capable model to date — and, unusually, its first users aren't developers or consumers. They're security teams.

The headline numbers (Google's reported results)

General capability

• #1 on the Artificial Analysis Intelligence Index v5.0, with a score of 59.
• #4 on WebDev Arena at 1768 Elo, up 89 points.
• Google says internal teams have run Carbon for days on long, multi-step engineering work, such as tracking down and patching subtle bugs in large codebases.

Security (measured without guardrails)

Google also says Carbon is more than twice as successful as Argon at writing secure code, and that Chrome's security team saw over 10x more high-quality patches than with another lab's large frontier model.

Why defenders go first

A model that's very good at finding and patching vulnerabilities is, by definition, also useful for exploiting them — note the 54.8% on ExploitGym. Google's answer is a staged release:

1. Now: a subset of cyber defenders get access.
2. Next: access widens to more trusted partners through Google DeepMind's Fairwind Program.
3. Later: general availability for developers, businesses and consumers is "coming soon," with no date.

Google says it's also participating in the U.S. government's voluntary pre-release model testing process. The logic: give the people patching systems a head start before the same capability is available to everyone.

Safety claims worth noting

Google reports several behavioral improvements versus Argon:

• 0.0% cheating on Terminal-Bench 4.0 (with a 0.4% rate of attempts).
• High-severity misalignment alerts in live coding environments down almost 80%.
• Unsupported claims roughly halved; "overreach" down 26% and "under-service" down 29%.
• 0.0% attack success on the Gray Swan indirect prompt-injection benchmark (Argon: 0.7%).

As with every lab's self-reported safety numbers, treat these as claims pending independent evaluation — but the prompt-injection result is notable given how much real-world agent risk comes from malicious instructions hidden in files and web pages.

What it will cost

Cached input is discounted 95% off the input price. The introductory rate matches Argon.

The bottom line

Carbon is Google's bid to retake the "best model" crown, and the security-first rollout is the more interesting part of the story. If staged access becomes the norm for frontier models with strong offensive capability, the gap between "announced" and "available" is going to get longer — and defenders who join programs like Fairwind will get months of advantage.