Google released Gemini 4 Carbon on Wednesday, a frontier model the company says delivers a 'significant leap' in cyberdefense. The model currently sits at the top of the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 and GPT-6 Astra.
Carbon is the second Gemini 4 release this month. Google shipped Argon two weeks earlier, and Carbon builds directly on that architecture with sharper coding, agentic task handling, and vulnerability discovery.
The stakes are timing as much as technology. Google spent most of 2026 trailing OpenAI and Anthropic in the frontier model race, and Carbon is the company's bid to reclaim the lead on a narrow but lucrative slice of the market: AI systems that can find real bugs before attackers do.
What Carbon Actually Scores
Google's announcement centers on two benchmarks. Carbon holds the number one position on SEC-bench Pro, which measures an AI model's ability to discover real-world software bugs. It also tops ExploitGym, which measures how well a model can turn a found vulnerability into a working exploit.
Those are not abstract wins. A model that can both find a bug and demonstrate how to exploit it is exactly what a defender wants to find first, and what an attacker wants to weaponize. Google is selling Carbon as a defense tool, and the company is restricting access accordingly: only a select group of cybersecurity defense companies can use it today. Other partners in Google's Fairwind Program will get access later.
The Artificial Analysis Intelligence Index v5.0 puts Carbon ahead of Claude Opus 5.5 and GPT-6 Astra overall. That composite ranking matters less to most buyers than the security-specific scores, but it signals that Google is competing on general capability as well as niche performance.
Built for Long, Messy Work
Google says Carbon is built for longer tasks. Internal teams have watched it run for days, persisting through dead ends to complete complex, multi-step engineering work. That claim is hard to verify from the outside, but it addresses a real weakness in earlier models: the tendency to declare a task done after hitting a shortcut.
Carbon is also better at resisting prompt injection attacks, according to Google. Prompt injection remains one of the least solved problems in deployed AI, so any improvement there has practical value even if the gains are incremental.
On honesty, Google says Carbon cuts 'unsupported claims' by two times compared with Argon. That is a measurable-sounding number attached to a vague category, and it is worth watching in independent testing. Models that sound confident while wrong have caused real damage in production, so a two-times reduction in unsupported claims, if it holds, would be more valuable than another benchmark point.
The Price Is the Story
Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. That is the detail security buyers will actually react to.
Frontier model pricing has been falling fast. Epoch AI's research puts the cost of achieving a given level of AI performance at roughly a 47 percent drop per quarter since 2023, or about 13 times per year. Google pricing Carbon at Argon's rate rather than charging a premium for the security capabilities suggests the company is buying market share, not margin.
It also means a cybersecurity team can run Carbon against large codebases without a budget shock. At $2 per million input tokens, scanning a million lines of code is cheap. The output side, at $10 per million tokens, is where costs accumulate if the model writes long reports or generates patches.
The Open Question
Google's claim that Carbon is better at finding vulnerabilities and generating secure code is plausible given the benchmark results. Whether it translates into fewer breaches in production is a different question, and the answer will not come from Google's own announcement.
The model is available to a select group of defense companies today, with Fairwind Program partners next. That limited rollout means independent verification will lag by weeks at minimum. Security teams should treat Google's numbers as a starting point, not a conclusion.
There is also the dual-use problem that Google does not fully address in its announcement. A model that tops ExploitGym, which measures exploit generation, is valuable to defenders and dangerous in the wrong hands. Google says Carbon is built for defense, but the capability is the capability. Restricting access to select partners is the current answer; it is not clear that answer scales as the model gets cheaper and more capable.
Google's Fairwind Program, which Carbon will expand through, is the company's channel for sharing advanced models with vetted partners. Fairwind partners get early access in exchange for feedback and responsible-use commitments. The program is Google's way of keeping sensitive capabilities out of general circulation while still getting them into production environments where they can be tested at scale. Whether that balance holds is an open question, because the history of restricted AI access is mostly a story of restrictions eroding once the model is copied, distilled, or reverse-engineered.
What Happens Next
Carbon's launch caps a busy month. Google released Argon two weeks ago, and Anthropic is reportedly preparing a new model of its own. OpenAI may launch a new model as well. The October model race is not over, and benchmark leadership has changed hands before.
For buyers, the practical advice is narrow. If you are a cybersecurity defense company with Fairwind access, Carbon is worth testing against your own codebase on your own metrics, not Google's. If you are not in that group, the pricing signal is the real news: frontier-grade security AI is available at $2 per million input tokens, and that price will pressure every other vendor in the market.
Google says Carbon reflects 'the pace at which we're pushing the frontier forward.' That is marketing. The real test is whether security teams find fewer bugs next quarter than they find today, and whether Google's restricted access model holds once the model is cheaper to run and easier to copy.
For more on the broader AI market, see the Crunchbase Q3 2026 funding report, and read our earlier coverage in AI category.