Google has made Gemini 4 Carbon, its most capable AI model, available to every paying developer, enterprise customer and top-tier consumer subscriber, two weeks after giving cybersecurity defenders an early look through a restricted access program.
In a post on Google's own blog, the company said Carbon is built for harder, longer-horizon tasks that benefit from spending more tokens to explore and verify answers, while its predecessor Argon remains "the ideal choice for the majority of software engineering and knowledge work." Carbon is now reaching developers through the Gemini API in Google AI Studio and Android Studio, enterprises through Gemini for business and Google Antigravity, and consumers who pay $100 a month for Google AI Ultra, via the Gemini app and AI Mode in Search.
A Clock on the Price
Google is charging an introductory $2 per million input tokens and $10 per million output tokens — matching Argon's rate — but that pricing expires November 30, after which the public-access model will cost $4 per million input tokens and $20 per million output tokens.
Carbon was first released on October 28 to a narrower group of cybersecurity defenders, including government agencies and critical-infrastructure operators, through what Google calls its Fairwind Program. According to TechRepublic's coverage of the staged rollout, Google cited Carbon's performance on benchmarks for finding, exploiting and patching software vulnerabilities as the reason those users went first; the company says a restricted version of the model found zero-day flaws in two major software vendors in under two hours as part of a program run with the security firm Wiz.
Google also reported that Carbon blocked 99% of simulated cyberattack sequences in internal testing, with a breach rate of 0.8% per attempt compared with 3.1% for Argon, and that external red-teamers at Trajectory Labs found no new universal jailbreaks across more than 13,000 requests. The company was more guarded about one weak spot: it said resistance to indirect prompt injection, where malicious instructions are smuggled into a model through documents or web content rather than typed directly, "remains an area for continued improvement."
