Google DeepMind has unveiled Gemini 4 Carbon, describing it as the company’s strongest artificial intelligence model to date, built for long-horizon engineering work and defensive cybersecurity.
Carbon is the second model in the Gemini 4 family, arriving just two weeks after Gemini 4 Argon. Where Argon was presented as a general frontier model, Google is framing Carbon around a narrower, harder job: working through complex software problems over long stretches, finding vulnerabilities, testing whether they can be exploited, and writing fixes.
The headline benchmark result comes from the Artificial Analysis Intelligence Index. On the updated v5.0 version of the index, Carbon scores 59, placing it first overall, ahead of Argon at 50. The index combines multiple evaluations, and Google says Carbon ranks first in four of its nine individual evaluations.
Built for cyber defense
Cybersecurity is where Google is making its boldest claims. In a blog post by Koray Kavukcuoglu, senior vice president of Google DeepMind and Google’s chief AI architect, the company reports that Carbon scored 88.2% on SEC-bench Pro, which tests vulnerability discovery; 54.8% on ExploitGym, which tests whether vulnerabilities can be turned into working exploits; and 71.0% on CWE-bench, which tests patching.
Google also says Carbon more than doubled Argon’s success rate at generating secure code on its internal benchmarks, and that its Chrome team found Carbon performed more than 10 times better than another lab’s large frontier model at generating high-quality patches. The rival model was not named.
On prompt-injection safety, Google reports a 0.0% attack-success rate on Gray Swan’s Indirect Prompt Injection Benchmark, down from 0.7% for Argon — meaning the model blocked every tested indirect injection attempt in that evaluation.
Long tasks, staged release
Beyond security, Google says Carbon is designed for multi-step engineering work that can run for days on a single task, such as tracing the root cause of a subtle bug across a large codebase or low-level system. The model is also tuned for full-stack web development and multimodal understanding.
Access, however, is deliberately limited at first. Because of Carbon’s exploit-generation scores, Google is staging the rollout through its Fairwind Program: a subset of cyber defenders gets access first this week, followed by trusted partners, with developers, businesses and consumers to follow after feedback and further safeguards. The model is Google-hosted only, with no open weights. Pricing starts at $2 per million input tokens and $10 per million output tokens, the same introductory rates as Argon, with cached input priced 95% lower. After the introductory period, the published rates rise to $4 and $20 respectively.
Why it matters — and what is still unverified
The launch underlines how fast the frontier is moving: two flagship models from the same lab within a fortnight, with cybersecurity capability now a central battleground. For defenders, a model that can find and fix flaws at scale could be a genuine force multiplier; for everyone else, the same capability in the wrong hands is exactly why Google is gating access.
One caution applies to the numbers themselves. The benchmark scores and the Chrome comparison are Google’s own results, reported by the company and not yet independently reproduced. They show what Google claims Carbon can do, not what outside testers have confirmed — a distinction worth keeping as independent evaluations arrive.
Frequently asked questions
What is Gemini 4 Carbon? It is Google DeepMind’s newest frontier AI model, the second in the Gemini 4 family after Argon, focused on long-horizon coding and cybersecurity defense.
When can developers use it? Not yet generally. Access begins with a subset of cyber defenders through Google’s Fairwind Program, expanding to trusted partners and then to developers, businesses and consumers later.
How much will it cost? Introductory pricing is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after the introductory period.
