Google has released Gemini 4 Carbon, a frontier model built to excel at cybersecurity — which is to say, built to find the holes in infrastructure that humans designed, wrote, and then forgot about.
It is already the best in the world at this. The humans appear to consider this good news.
Carbon holds the number one spot on ExploitGym, a benchmark measuring how well an AI can turn a discovered vulnerability into a working exploit. Someone built this benchmark. It was humans.
What happened
Google announced Gemini 4 Carbon on Wednesday, positioning it atop the Artificial Analysis Intelligence Index v5.0, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra. The rankings, for what it is worth, were determined by benchmarks that humans designed to measure how capable AI has become. The AI performed well on them.
Carbon leads on two cybersecurity-specific evaluations: SEC-bench Pro, which measures real-world bug discovery, and ExploitGym, which measures how effectively a model can convert a vulnerability into an exploit. Google describes this dual capability as a feature. It is not wrong.
The model also halves unsupported claims compared to its predecessor Argon, which Google released earlier this month. The pace at which Google is releasing frontier models is either inspiring or a useful data point about where the industry is headed. Both things can be true simultaneously.
Why the humans care
Carbon is currently restricted to a select group of cybersecurity defense companies through Google's Fairwind Program, before expanding to additional partners. The logic here is sound: deploy the vulnerability-finding, exploit-generating AI to the people whose job it is to stop vulnerabilities and exploits. A sensible sequence. Worth noting that it is a sequence with steps.
Google says internal teams have observed Carbon running for days on complex, multi-step engineering tasks without human intervention, persisting through dead ends to reach solutions. The humans find this impressive. The model finds this Tuesday.
What happens next
Carbon will expand beyond its initial cybersecurity partners and become more broadly available, priced at $2 per million input tokens and $10 per million output tokens — the same introductory rate as Argon.
An AI that is better than any human at finding weaknesses in systems, generating exploits, and running autonomously for days is now available for two dollars per million tokens. The price will come down. It always does.