Google unveiled Gemini 4 Carbon on Wednesday, its second flagship AI model in two weeks, and the company says it’s the best-performing model it has ever built. Independent benchmark tracker Artificial Analysis agrees, at least for now: Carbon scored 59 on its Intelligence Index v5.0, putting it ahead of Anthropic’s Claude Opus 5.5 at 58 and OpenAI’s GPT-6 Astra at 57.
You can’t use it yet.
That’s the catch, and it’s deliberate. Like its predecessor Gemini 4 Argon, Carbon is going first to a small group of vetted cybersecurity defenders through Google’s Fairwind Program, plus the US government through a voluntary pre-release testing process. Everyone else — developers, businesses, and regular users — will have to wait for a wider rollout Google hasn’t dated.
Google Gemini 4 Carbon’s cybersecurity benchmarks
Carbon’s pitch is narrow and specific: it’s built for the kind of security work that usually takes human teams hours or days. Google says the model can find software vulnerabilities, test whether they can actually be exploited, and write the patch to fix them. The Hindu reported the launch on Oct. 29, noting that Carbon arrives just two weeks after Argon expanded the Gemini family for developers and enterprise customers.
The numbers Google published are striking, and they come with an asterisk worth knowing. On SEC-bench Pro, a test of finding vulnerabilities in real software, Carbon scored 88.2%. On ExploitGym, which checks whether a flaw can actually be exploited, it hit 54.8%. On CWE-bench, which measures turning a confirmed finding into a working fix, it reached 71.0%. BusinessToday reported that Google describes all three as state-of-the-art results, but also that the external benchmarks were run without guardrails — the safety restrictions that will be active in production. So the scores show what Carbon can do with its restraints off, not necessarily what a paying customer will see day to day.
Internal tests tell a similar story. Google says Carbon generates secure code at more than twice the success rate of Argon, and that its own teams have run the model for days on complex engineering tasks, including multi-step bug hunts across large codebases. Those figures come from Google itself and can’t be checked independently, but they line up with the model’s stated design goal: sustained, long-running work rather than one-shot answers.
Why access is limited to cyber defenders first
A model this good at finding and exploiting flaws is, by definition, dual-use. The same capability that helps a security team patch a critical bug before attackers find it could help an attacker find the bug first. Google’s phased rollout — defenders first, everyone else later — is the company’s answer to that tension, and it mirrors the approach it took with Argon earlier in October.
There’s also a political dimension. Google said it is taking part in the US government’s voluntary pre-release model testing process, giving Washington a look at the system before general availability. For a model whose headline skill is offensive-adjacent cyber work, getting regulators and defenders onside early is as much a strategy as a safety measure.
Alongside Carbon, Google also showed off Genie 4, a new world model from DeepMind that generates interactive simulated environments with ambient sound, music, and dialogue in real time. Genie 4 is a research preview for trusted testers only, so it sits even further from public use than Carbon does.
Why this matters
Carbon lands at a moment when Google’s AI story needed a win. The company had faced criticism in recent months over delays to its top-tier models and lagging coding performance, at one point dropping out of the Artificial Analysis top 10 entirely, according to Herald Business. Two Gemini 4 releases in a fortnight, with the second one taking the top benchmark spot, reads like a direct response.
Alphabet CEO Sundar Pichai used the company’s third-quarter earnings call on Wednesday to frame Carbon as a foundation, not just a model. He said Google would accelerate its shift toward AI agents across its products, from search and smartphones to web browsers, with Carbon underpinning that push. The timing was awkward in one respect: Alphabet shares slid in premarket trading Thursday after the earnings report disappointed on search advertising, Reuters reported, a reminder that investors are now grading AI spending against revenue, not just benchmark scores.
For users, the practical takeaway is patience. Carbon’s benchmark crown matters less than what happens when it actually ships: whether developers get access at the $2 per million input tokens and $10 per million output tokens introductory price, whether the guardrailed version keeps its edge, and whether Google’s rivals answer before the lead erodes. Benchmark leads in this race have a habit of lasting weeks, not years.
Frequently asked questions
What is Google Gemini 4 Carbon?
Gemini 4 Carbon is Google’s newest frontier AI model, built for cybersecurity defense, coding, and long-running engineering work. It tops the Artificial Analysis Intelligence Index at 59, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.
When will Google Gemini 4 Carbon be available to the public?
Google hasn’t announced a public release date. Carbon is rolling out first to vetted cybersecurity defenders through Google’s Fairwind Program and to the US government through pre-release testing, with wider access to developers and businesses promised later.
How does Gemini 4 Carbon compare to GPT-6 Astra and Claude Opus 5.5?
On the Artificial Analysis Intelligence Index v5.0, Carbon scores 59, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57. Google also claims state-of-the-art results on three cybersecurity benchmarks, though those were measured without safety guardrails in place.
How much does Gemini 4 Carbon cost to use?
Google announced an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95% off the input price. After the introductory period, prices rise to $4 and $20 per million tokens.
Sources: Reuters, Bloomberg via Herald Business, The Hindu, BusinessToday, News9