Instant Champion: The Most Powerful Model You Can’t Use
Google has surprise-released its latest frontier AI model, Gemini 4 Carbon, delivering what the company calls “significant gains” in coding, agentic tasks, and cybersecurity. It debuts at the top of the Artificial Analysis Intelligence Index (v5.0), leaving Claude Opus 5.5 and GPT-6 Astra trailing behind — yet most people won’t get to touch it anytime soon: Carbon is currently available only to a select group of cybersecurity defense companies, with access later expanding to other partners in its Fairwind Program.
A “Significant Leap” in Cyberdefense
Google says Carbon achieves a significant leap in cyberdefense over Argon, released earlier this month, making it better at finding vulnerabilities and generating secure code. On SEC-bench Pro — a benchmark measuring an AI model’s ability to discover real-world bugs — Carbon currently holds the number one spot. It also dominates ExploitGym, which measures how effectively a model can turn a vulnerability into working exploits. In other words, this model is fluent on both sides of the security fence — precisely the capability at the heart of the industry’s ongoing debate over frontier AI risk.
Marathon Performance: Running for Days Without Giving Up
Carbon’s other headline feature is long-horizon persistence. Google says its internal teams have watched the model “run for days to persist through dead ends to tackle complex, multi-step engineering tasks.” It is also better at avoiding shortcuts, resisting prompt injection attacks, and collaborating with users. On honesty, Google claims Carbon cuts “unsupported claims” by two times compared to Argon, boosting transparency and reliability in tandem.
Aggressive Pricing: From $2 Per Million Tokens
Carbon launches at the same competitive introductory price as Argon: $2 per million input tokens and $10 per million output tokens. Against similarly positioned rivals, that price is strikingly aggressive — a signal that Google intends to pair “stronger and cheaper” to grab share in the enterprise AI market.
Why Can’t You Use It? The Fairwind Logic
“Strongest model, defenders first” is not marketing spin — it’s risk management. When a model is powerful enough to turn vulnerabilities into working exploits (top-of-chart ExploitGym), unlimited public release would amount to handing out attack tooling. Google is instead letting cybersecurity defense firms validate and deploy Carbon first, then gradually widening access through the Fairwind Program — consistent with the industry’s recent discussions around restricting “critical cybersecurity-tier” models.
Conclusion: The Frontier Race Enters “More Power, More Caution”
Gemini 4 Carbon marks a new phase of the frontier AI race: model capability keeps climbing, but access to top-tier abilities is being gated more carefully. Three things to watch: first, when Carbon reaches broader availability through Fairwind; second, whether “stronger at half-price-class pricing” forces rivals to cut prices; and third, whether cybersecurity benchmarks like SEC-bench Pro and ExploitGym become the new standard for evaluating frontier models. One thing is certain — the competition has expanded from “who is smarter” to “who can better guard the world’s digital doors.”
