Google opens Gemini 4 Carbon to paying developers and Ultra users Google opens Gemini 4 Carbon to paying developers and Ultra users

An access change, not a second unveiling

Google first announced Carbon on October 28. That release described staged access for cybersecurity defenders through the Fairwind Program and engagement with government pre-release testing. Developers and consumers were still waiting. The earlier post positioned Carbon around longer engineering tasks, including investigation and repair across large codebases. It also said the model used the safety stack developed for Argon. Today’s developer announcement supplies the practical change: the paid API tier in AI Studio and Android Studio is available, while Ultra subscribers can try Carbon in AI Studio and Antigravity. A team evaluating coding agents can therefore move from watching demonstrations to testing its own work.

The price has a calendar attached

The public model starts at two dollars per million input tokens and ten dollars per million output tokens. Google’s footnote sets November 30 as the end of that offer. The subsequent rates are four and twenty dollars respectively. For an illustrative job consuming one million input tokens and 100,000 output tokens, uncached token charges total three dollars initially. At the later rates, the same quantities cost six dollars. These calculations exclude tools, storage and other charges. They isolate the effect of the announced rate change. Actual agent costs also depend on how much work the model repeats before it finishes.

More workers still need a coordinator

Google is also previewing a swarm feature in Antigravity. It assigns a prompt to coordinated agents across isolated workspaces, with separate verification roles. Carbon can handle planning and auditing alongside lighter models. This changes the unit of comparison from one answer to a completed project. Parallel workers can investigate different parts of a problem, while integration and acceptance remain separate jobs. A fast first draft and a verified final change are different outputs. That distinction becomes more important when several agents contribute to the same deliverable. Antigravity’s own post warns that parallel teams can consume quota much faster than standard chat, even when idle agents use no tokens.

A new test measures finished files

A separate November 10 announcement introduced the Multimodal Knowledge Agent benchmark with Artificial Analysis. Its 200 evaluation tasks span ten fields and 83 occupations, using documents whose visual structure matters. Models run in the same open-source Stirrup harness with matching tools and budgets. The rubric weights factual accuracy at 75 percent and visual design at 25 percent. Google reports Carbon placed second for quality, at roughly one-sixth the average task cost of the highest-scoring model. The evaluation set remains private, while ten separate example tasks are public. That design offers a more concrete comparison than a polished demonstration alone: it examines the delivered files against explicit requirements.