Google expanded access to Gemini 4 Carbon on November 10, only two weeks after introducing its latest frontier model with limited availability for cybersecurity defenders. The change turns that initial announcement into a broader release: Carbon is now available through the API to paying developers and enterprises in Google AI Studio and Gemini for business, and it is rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search.[1]

The company set introductory pricing at $2 per million input tokens and $10 per million output tokens, matching Gemini 4 Argon. Google did not specify in the announcement when that rate will end. The company still recommends Argon for most software-engineering work, while positioning Carbon for tasks that demand deeper reasoning, dense-document analysis, scientific research, coding, and long-horizon agent workflows.[1]

The most visible product addition is `/swarm`, a preview feature inside Google Antigravity. From a single prompt, the system organizes multiple agents in isolated workspaces and lets users monitor their progress. Google suggests assigning Carbon roles such as architecture, coordination, and auditing while faster models handle parallel work. Company examples include three-dimensional reconstructions built from photographs, plans, and documents, but these remain demonstrations selected by Google rather than independent tests of routine performance.[1]

For coding and agent work, Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench. CursorBench provides a more concrete public check: the “Gemini 4 Carbon High” configuration records 60.4% on ambiguous, multi-file tasks, narrowly ahead of 60.0% for “Fable 5.2 Max.” The benchmark itself warns that results are subject to variance and small score differences may not be statistically meaningful.[2] The ranking therefore should not be treated as proof of universal model superiority.

Carbon also retains an unusual emphasis on cybersecurity. Google says that, when accessed through its restricted Fairwind Program, the model leads evaluations for finding, exploiting, and patching vulnerabilities. The company reports that its internal Project Fortify system reached 100% recall in scanning and 95% precision in triage on an internal benchmark, as well as a relative improvement of more than 25% over Argon in open-source repository tests. These are figures from Google and early partners; they are not equivalent to a comprehensive external audit, and they do not mean offensive capabilities are openly available without restrictions.[1]

An independent AI Weekly brief matches the launch’s central details —paid access, introductory pricing, benchmark results, and the `/swarm` preview—although it points to Google’s announcement as the original source.[3] The strongest currently available evidence confirms availability, pricing, and Cursor’s published ranking. Broader claims about productivity, security, and organizational performance still require independent validation.

Carbon’s wider release matters less because of any single score than because it combines a high-reasoning model with a native multi-agent interface. Google is moving that architecture from restricted testing into products used by developers, enterprises, and consumers. The next test will be practical: how much control users retain over each agent, how actions are logged, and whether benefits shown in demonstrations survive real, prolonged tasks whose outcomes are harder to verify.

Sources

[1] Read More — Gemini 4 Carbon is now available — Google
[2] Read More — CursorBench 4.0 — Cursor
[3] Read More — Google Ships Gemini 4 Carbon to Paid API and AI Ultra — AI Weekly