Gemini 4 Carbon puts Google at the top of the leaderboard, with its new cyber capabilities restricted to defense partners

Google has released Gemini 4 Carbon, its new frontier model. According to the company's announcement, reported by The Verge, it brings "significant gains" in coding, agentic tasks and cybersecurity. Google says Carbon ranks first on the Artificial Analysis Intelligence Index (v5.0), a composite score that aggregates many tests, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra.

On security, Google describes a "significant leap." Carbon leads SEC-bench Pro, which measures a model's ability to find real-world bugs, and ExploitGym, which measures how well it turns a vulnerability into a working exploit. Google says internal teams have seen it "run for days" through complex engineering tasks without giving up at dead ends. The company also claims it resists prompt injection better (attacks that slip malicious instructions into the text a model reads) and that it halves "unsupported claims" compared with Argon, the model Google released earlier this month. Pricing stays at Argon's level: $2 per million input tokens and $10 per million output tokens.

The key detail is in the deployment. Carbon is currently available only to a select group of cyber defense companies and will later expand to other partners in Google's Fairwind Program. In practice, Google is treating a model that is good at writing exploits as a dual-use capability that has to be rolled out in stages. That is the right instinct. It follows the precedent of models being handled as strategic assets rather than ordinary products.

Our thesis: this is a real move up for Google, and it challenges our working assumption that Google sat behind Anthropic and OpenAI on the hardest tests. But caution is warranted. The rankings come from Google's own announcement. Leading a composite index or a specific benchmark is not the same as leading in everyday use. Version and configuration changes have shifted these scoreboards before. For now, we're treating Carbon's lead as a claim to verify, not settled fact, until independent evaluations and unguided expert tasks confirm it.

Our reading: the frontier race now runs on two tracks at once. The first is general capability, where the gap between the top three labs is narrowing to a matter of weeks. The second is how responsibly they release what they build. Two Gemini models in one month at a competitive price will push margins down across the market, and that is good news for wider access. The fact that the cyber capability starts with defenders is good news too. In the short term, the risk of tools that can generate exploits is real. In the long term, models that can find and patch vulnerabilities faster than attackers can exploit them are one of the clearest ways AI can make digital infrastructure safer for everyone.