Google Releases Gemini 4 Carbon to Public, Previews ‘Swarm’ Agent Feature
Google opens its strongest model, Gemini 4 Carbon, to paying users and introduces a ‘swarm’ feature that turns single prompts into coordinated agent teams.
U.S. tech giant Google officially released its latest artificial intelligence model, Gemini 4 Carbon, to the general public on November 10. For the previous two weeks, access was restricted to a select group of “cybersecurity defenders.” Google claims Carbon is its “strongest model” yet, capable of competing with the best systems from Anthropic and OpenAI.
According to The Verge, Google stated that Gemini 4 Carbon “pushes the frontier even further” than Argon, another frontier model launched by the company just last month. While Argon remains the ideal choice for most software engineering knowledge work, Carbon is built for tasks that test the limits of a model’s intelligence, such as complex, multi-step coding and open-ended scientific research.
To highlight Carbon’s utility for long-horizon work involving AI agents, Google announced it is also previewing a “swarm” feature on its agentic development platform Antigravity. This feature allows users to turn a single prompt into a “coordinated team of agents,” enhancing capabilities for handling complex tasks.
Carbon is now rolling out via the Gemini API in Google AI Studio and Gemini for business for all paying developers and enterprises. It is also available to all Google AI Ultra subscribers in the Gemini app, the AI Mode in Search, and on Antigravity. It is not clear when—or if—Carbon will be available to lower tier subscribers or free users.
In terms of pricing, Google continues its highly competitive strategy. Carbon launches with an introductory price of $2 per million input tokens and $10 per million output tokens, roughly one-fifth the cost of OpenAI’s GPT-6 Astra and one-quarter the cost of Anthropic’s Fable 5.2 model. Even after these prices double at the end of November, Carbon will still be significantly cheaper than frontier rivals on a per-token basis, though this does not necessarily translate to lower total costs as models can use different numbers of tokens for the same tasks.