Google's New Gemini 4 Carbon Beats GPT-6 and Claude on Key Benchmarks, With a Focus on Hacking Defense

Google released its newest frontier AI model on Wednesday, and it's not aimed at chatbot users or developers building apps. It's aimed at hackers, or rather, the people trying to stop them.

Gemini 4 Carbon is Google's latest large language model, and according to the company's announcement, it delivers "significant gains" in coding, agentic tasks, and cybersecurity. That last category is where Carbon is making its loudest claims. Google says the model represents a "significant leap" in cyberdefense, with real improvements in finding software vulnerabilities and writing secure code.

The timing matters here. Carbon arrives just weeks after Google released Argon, another frontier model, earlier this month. That's a fast release cadence even by AI industry standards, and Google is framing it as evidence of how quickly it's moving. In the company's own words, Carbon reflects "the pace at which we're pushing the frontier forward."

Topping the charts, at least for now

Google is touting Carbon's performance on the Artificial Analysis Intelligence Index (v5.0), a widely watched benchmark that ranks AI models on general capability. According to Google, Carbon currently sits in the top spot, ahead of Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra. For a field where leadership claims shift every few weeks depending on whose model just shipped, that's a notable, if likely temporary, bragging right.

The more specific claims are in security-focused benchmarks that most casual AI users have never heard of. Carbon is ranked number one on SEC-bench Pro, which tests how well a model can discover real-world software bugs, and on ExploitGym, which measures a model's ability to take a known vulnerability and turn it into a working exploit. Those are not benchmarks about writing better emails or summarizing documents. They're about whether an AI can do the kind of offensive and defensive security work that currently requires skilled, expensive human researchers.

That framing is deliberate. Google isn't pitching Carbon as a general-purpose assistant for the masses right now. It's positioning the model as a tool for the security industry specifically, and it's rolling it out accordingly.

Access is limited, for now

Gemini 4 Carbon isn't available to the general public or even to Google's broader developer base yet. According to the announcement, it's currently restricted to a select group of cybersecurity defense companies. Google says access will eventually widen to other partners through its Fairwind Program, though the company didn't specify a timeline for that expansion.

That controlled rollout makes sense given what the model is reportedly capable of. A system that's unusually good at finding exploitable bugs is valuable to defenders trying to patch systems before attackers find the same holes. It's also, in the wrong hands, a tool that could speed up attacks. Starting with vetted security partners rather than a wide release suggests Google is being deliberate about who gets early access to that capability.

Built to grind through long jobs

Beyond the security benchmarks, Google is emphasizing Carbon's stamina. The company says its internal teams have watched the model work on complex, multi-step engineering problems for days at a time, persisting through dead ends rather than giving up or looping on the same failed approach. That's a meaningful shift in how these models are being evaluated. For much of the generative AI boom, the headline metric was how smart a model sounded in a single response. Increasingly, the metric that matters for real engineering and security work is whether a model can stick with a hard, open-ended task over an extended stretch without a human steering every step.

Google also says Carbon is better at avoiding shortcuts, a known failure mode where models technically complete a task by gaming the instructions rather than genuinely solving the problem. The company says the model is more resistant to prompt injection attacks, a security concern specific to AI systems where malicious instructions get smuggled into content the model processes, tricking it into acting against its actual user's interests. For a model being marketed partly as a security tool, resisting that kind of manipulation isn't a nice-to-have. It's close to a prerequisite.

Google is also claiming gains on honesty. The company says Carbon cuts "unsupported claims" by half compared to Argon, its predecessor from earlier this month. That's a direct nod to one of the most persistent criticisms of large language models: their tendency to state things confidently that aren't actually true, often called hallucination. Halving that rate, if it holds up under outside scrutiny, would be a real improvement rather than a marginal one.

Pricing stays put

Despite the upgrades, Google isn't charging more for Carbon than it did for Argon. The model launches at the same introductory pricing: $2 per million input tokens and $10 per million output tokens. Tokens are the chunks of text models process and generate, and pricing per million tokens is the standard way AI labs bill for API access. Holding the price steady while claiming benchmark leadership is Google's way of signaling that it intends to compete on capability without asking customers to pay a premium for it, at least for now.

What happens next is the real test. Benchmark rankings in this industry have a short shelf life, and Carbon's claimed lead over GPT-6 Astra and Claude Opus 5.5 will likely face a challenge from a rival lab within weeks. The more interesting question is how Carbon performs once it's in the hands of actual cybersecurity defense companies, and how quickly Google expands access through the Fairwind Program beyond that initial group. A model that's genuinely better at finding and fixing real vulnerabilities would matter a lot more than another benchmark chart-topper.