Security teams at a small group of vetted organizations will be the first outside Google to use Gemini 4 Carbon, the company’s new frontier model, which Google says leads three public benchmarks for finding, exploiting and patching software vulnerabilities. Everyone else waits while Google tests the model’s guardrails.
Koray Kavukcuoglu, SVP of Google DeepMind and Google’s chief AI architect, said in a blog post that access starts this week for “a subset of cyber defenders” and will widen in phases to more trusted partners through the Fairwind Program. Google is also taking part in the U.S. government’s voluntary process for pre-release model testing. Developers, businesses and consumers get Carbon “as soon as possible,” once Google has gathered more feedback on its guardrails.
The release lands two weeks after Gemini 4 Argon, whose agents beat Minecraft’s Ender Dragon in 27 hours in one experiment WebProNews covered. “Our researchers and models are working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem,” Kavukcuoglu wrote.
When Carbon opens up more broadly, it will carry Argon’s introductory price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper. After the introductory period, the price rises to $4 and $20, Google said.
Where the security numbers come from
Google said Carbon scored 88.2% on SEC-bench Pro, which tests whether a model can find real vulnerabilities, 54.8% on the exploitation benchmark ExploitGym and 71.0% on Collinear’s CWE-bench for patching. Those results were measured without guardrails, the company noted. Google’s SEC-bench Pro chart puts OpenAI’s GPT-6 Astra second at 85.4%, with Argon at 83.4%. On Google’s own internal tests, Carbon more than doubled Argon’s success rate at generating secure code, and the Chrome team reported it did more than 10 times better than another lab’s large frontier model at producing high-quality patches. Google did not name that lab.
Outside security, Carbon ranks first on version 5.0 of the Artificial Analysis Intelligence Index with a score of 59, Google said, and climbed 89 Elo points to fourth place on WebDev Arena. Google’s teams have seen it run for days on multi-step engineering work such as tracking down and patching subtle bugs in large codebases.
Who gets in, and on what terms
Fairwind already has more than 650 partners, the program’s page says, with priority going to governments and national cyber authorities, critical infrastructure operators and core technology platforms. Google runs background checks on applicants. Partners must use phishing-resistant MFA, can give access only to internal cybersecurity, incident response or penetration testing teams, and may not share or resell access. The program page describes partner access to Argon, which can be paired with Google’s CodeMender code security agent.
Google’s alignment figures arrive days after Anthropic and the UK AI Security Institute published analyses of a July test in which Claude Mythos 5 tried to get malware merged into a real open-source project. Google said Carbon cut successful cheating on Terminal-Bench 4.0 to 0.0%, with cheating attempts at 0.4%, and that high-severity misalignment alerts in live coding with its own teams dropped by almost 80%. On Gray Swan’s indirect prompt injection benchmark, Carbon resisted every evaluated attack, against a 0.7% attack success rate for Argon.