Originally published on The Daily Flare.
Google began rolling out Gemini 4 Carbon on November 10, 2026, to paying developers and enterprises, while previewing a feature that coordinates multiple AI agents in Antigravity.
The company describes Carbon as its strongest model, built for demanding work such as complex, multi-step coding, debugging and open-ended scientific research. Google says its Gemini 4 Argon model, launched last month, remains the better fit for most software engineering and knowledge-work tasks. Carbon is designed to use more tokens than Argon to explore problems, reason through them and run extended verification loops.
Access is being extended through the Gemini API in Google AI Studio and Gemini for business for paying developers and enterprises. Google AI Ultra subscribers can use Carbon in the Gemini app, AI Mode in Search and Antigravity. Google also says it is providing 10,000 scientists with six months of free access to the model for research.
Alongside the rollout, Google is previewing the “swarm” feature on Antigravity, its agentic development platform. A user can turn one prompt into a coordinated team of agents and follow their work in real time. Google says Carbon can take on roles requiring deeper reasoning, including architectural design, coordination and auditing, while faster, lighter models work in parallel spaces and adversarial verifiers check changes.
Google’s announcement gives examples of projects made with the feature. In one, a team used historical photographs and engineering plates to create an interactive Three.js exhibition about the Eiffel Tower’s construction from 1887 to 1889. In another, agents used seven public city council planning PDFs, containing 206 extracted engineering drawing views, to build a 906,000-triangle interactive reconstruction of Google’s Platform 37 office over 10 overnight loops.
TestingCatalog, which tried an experimental Swarm command in Antigravity, observed a dashboard with a shared message board, agent roles and a graph showing work and messages. Its test launched three agents, including ones named Greeter and Verifier. TestingCatalog also reported a warning that the feature could consume substantial quota, a consideration when agents use token-priced models.
Google says Carbon ranked first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than twice lower cost per task than the next-best external model. On the Multimodal Knowledge Agent benchmark, developed with Artificial Analysis to test agents working with dense files containing items such as charts, images and tables, Google says Carbon ranked second and had an average cost per task more than six times lower than the leading model.
Those results are company-cited benchmark claims. The Verge noted that companies can select benchmark results that present their systems favourably, and that scores do not necessarily establish performance across every real-world task.
Carbon is launching at an introductory price of $2 per million input tokens and $10 per million output tokens, matching Argon’s introductory rates, Google said. The Verge reported that the rates are due to double at the end of November, after the introductory period, and that even then Carbon would remain cheaper per token than the frontier competitors it compared. The outlet also cautioned that token prices alone do not determine the cost of a task, because models may use different numbers of tokens to complete the same work.
Top comments (0)