Google Opens Gemini 4 Carbon to Developers: $2 per Million Tokens and a Swarm of Agents

Google has released Gemini 4 Carbon to the general public. For two weeks the model was available only to a narrow circle of "cybersecurity defenders"; now all paying developers and companies have it. It is Google's second flagship model in a month: Argon only came out in October. For anyone choosing a model for a product, three things matter: what Carbon is for, what it really costs, and what the "swarm" of agents in Antigravity is.

Carbon and Argon: which one to pick

Google does not say "Carbon is best at everything." According to the company, Argon remains "the ideal choice for most software development and knowledge work tasks." Carbon is for cases where you hit the ceiling of the model's intelligence: complex multi-step coding and open-ended scientific research.

In practice, this resembles the familiar split between a fast workhorse and heavy artillery. For autocomplete, reviewing small pull requests, summarizing documents, and typical business scenarios, Argon will most likely be enough. Carbon is worth bringing in when an agent has to work on a task for hours, keep dozens of steps in mind, and not lose the thread.

Availability is as follows: through the Gemini API in Google AI Studio and Gemini for business for all paying developers and companies. Google AI Ultra subscribers will get the model in the Gemini app, in AI Mode in Search, and in Antigravity. Whether Carbon will come to cheaper subscriptions and free users is unknown.

A "swarm" of agents in Antigravity

Alongside the model, Google showed a preview of the swarm feature in its agentic development platform, Antigravity. The idea is that a single prompt turns into a "coordinated team of agents." Instead of one long conversation with the model, you describe a goal, and the platform breaks it down into roles and subtasks on its own.

The announcement offers few details, so it is too early to judge the quality of the coordination. But the purpose is clear. Long tasks, such as migrating a large repository or going through a set of scientific papers, fit poorly into a single context. Several agents with different subtasks can work in parallel. This also raises the risk that one agent's error will spread to the others. For now it is a preview, and production processes should not be built on it.

Price: cheap, but with caveats

Carbon launches with an introductory price of $2 per million input tokens and $10 per million output tokens. By The Verge's estimate, that is roughly five times cheaper than OpenAI's GPT-6 Astra and four times cheaper than Anthropic's Fable 5.2, which came out yesterday. At the end of November prices will double, to $4 and $20, but even then Carbon will remain noticeably cheaper than competitors per token.

A caveat matters here: price per token is not price per task. Different models spend different numbers of tokens on the same job. If one model "thinks" twice as long, all the savings can vanish. So run your real scenarios on Carbon and compare the final cost of a solved task, not the price list.

What about benchmarks

Google claims that Carbon competes with the best systems from Anthropic and OpenAI. The benchmarks the company cites place it among the strongest models on the market, but on most metrics it still trails Fable 5.2. As a reminder, companies pick results that favor them for publication, so the figures in the announcement should be treated with caution.

What to do

• If you already have agentic pipelines on Argon, don't migrate wholesale. Test Carbon only on your hardest and longest tasks.
• Budget for the price doubling from the end of November and calculate cost per task.
• Don't tie critical processes to swarm while the feature is in preview.
• Compare Carbon with Fable 5.2 on your own data: by Google's benchmarks it is weaker, but noticeably cheaper.

Something else matters more for the market. For several months the frontier looked like a two-horse race between Anthropic and OpenAI. Now Google is back in that race, and with an aggressive price. For developers, that means more choice and more leverage in price negotiations.