Google on Wednesday, Oct. 28, announced Gemini 4 Carbon, which it calls its next frontier model, the term for the most capable AI models, ones that match or beat the best that exist today. Google says Carbon is its strongest model yet, with the best results so far at computer security and at long, many-step work. It arrives two weeks after Google launched Gemini 4 Argon, the model it builds on.
Few people outside Google can use it yet. Starting this week, Google is giving Carbon to some cyber defenders, and then, gradually, to more trusted partners through its Fairwind Program. That program gives early access to groups such as governments, healthcare providers and telecom companies, and Google says it has more than 650 partners. Google says it is taking part in a voluntary U.S. government process for testing models before release, and will open Carbon to developers, businesses and the public as soon as possible, once it has gathered feedback and improved its safety limits. It has not given a date.
Security is the main pitch. On a test of finding real vulnerabilities, weaknesses in software that an attacker could use, Google says its own runs put Carbon at 88.2 percent, ahead of the 85.4 percent OpenAI reported for its GPT-6 Astra. On a test of whether a model can turn a flaw into a working attack, a check defenders increasingly ask AI models to make when deciding which flaws to fix first, Google says Carbon set a new best, measuring it with both the standard six hours per task and a longer ten-hour limit. Google ran its own tests with Carbon's safety limits switched off, while the rivals' scores come from their own reports, so the results may not be measured the same way. Google also says its Chrome team found Carbon more than 10 times better than a large model from another company at writing good fixes for security flaws, without naming that model. Google's page for that early-access program warns that the same skills that help defenders find and fix flaws could help attackers just as much in the wrong hands.
Carbon does not lead everywhere. Google's own comparison puts it behind both GPT-6 Astra and Claude Opus 5.5 on a test of hard programming tasks, 61.2 percent against 65.5 and 62.3 percent, and behind Claude Opus 5.5 on a test of using a computer's apps on its own, 81.5 percent to 87.2 percent. Google says Carbon ranks first on an overall index from Artificial Analysis, a company that compares AI models. Vals AI, an independent company that tests AI models on finance, legal, tax and coding work, ranks Carbon second of 29 models on its overall index at 70.36 percent, just behind Anthropic's Claude Sonnet 5.5 at 70.85 percent, a model Google's comparison leaves out. Claude Opus 5.5 follows at 70.18 percent, so the three are within less than a point of each other, closer than Vals's own margin of error for Carbon's score.
Google says Carbon can keep working for days on hard engineering problems, such as tracking down and fixing subtle bugs in large bodies of code. It says it also trained Carbon to take fewer shortcuts: on one coding test, it rarely tried to game the test instead of doing the task, and none of those attempts succeeded. In Google's internal testing, it made about half as many unsupported claims as Argon. Those are Google's own measurements.
Prices are set per million tokens, the small chunks of text, often parts of words, that AI models read and write. When Carbon opens up, developers will pay the same introductory price as for Argon: $2 per million tokens the model reads and $10 per million it writes, rising later to $4 and $20. Google has not said how long the introductory price will last.