AI
Google launches Gemini 4 Carbon and previews agent clusters
Google is rolling out Gemini 4 Carbon out to paying developers and enterprises, saying it is designed for difficult tasks such as complex multi-step programming and open-ended scientific research. It is also previewing a “swarm” feature on the Antigravity platform that can turn a single prompt into a collaborative team of agents. Entry pricing is lower than leading competitors.
Two weeks after opening it to some “cybersecurity defenders,” Google has officially launched its latest AI model, Gemini 4 Carbon, to the public. Google says Carbon is its “most capable model to date,” going further than the frontier model Argon released last month and able to compete with top systems from Anthropic and OpenAI.
Google says Argon may still be the ideal choice for most software engineering and knowledge work, while Carbon is built for tasks that test the limits of model intelligence, such as complex multi-step programming and open-ended scientific research. To demonstrate its usefulness in long-horizon agent tasks, Google is also previewing the “swarm” feature on its agent development platform Antigravity, allowing users to turn one prompt into a coordinated team of agents.
Carbon is available to all paying developers and enterprises through the Gemini API in Google AI Studio and Gemini for Business, as well as to Google AI Ultra subscribers in the Gemini app, Search AI Mode, and Antigravity. It remains unclear when or whether users on lower-tier subscriptions or free tiers will gain access.
The entry price is $2 per million input tokens and $10 per million output tokens. According to Google, that is roughly one-fifth the cost of OpenAI’s GPT-6 Astra and one-quarter the cost of Fable 5.2, which Anthropic released the previous day. Prices will double after the promotional period ends at the end of November, but on a per-token basis they will still be significantly below frontier competitors. Reports also caution that different models may require different numbers of tokens for the same task, and vendors may selectively present benchmark results.