Google rolls out Gemini 4 Carbon with agent swarm feature to paying users

Gemini 4 Carbon reaches the public

Google has begun rolling out Gemini 4 Carbon, which it calls its strongest AI model to date. According to The Verge, the public release comes two weeks after the model was first given to a select group of cybersecurity defenders, and Google positions it as a system that can compete with the best offerings from Anthropic and OpenAI.

The model is available now through the Gemini API in Google AI Studio and in Gemini for Business for all paying developers and enterprises, The Verge reports. It is also reaching Google AI Ultra subscribers in the Gemini app, in Search's AI Mode, and on Antigravity, Google's agentic development platform. There is no word yet on when, or whether, lower-tier subscribers or free users will get access.

A second frontier model in two months

Carbon arrives just a month after Argon, the other frontier model in Google's lineup. The company says Carbon pushes the frontier further than Argon, but it still expects Argon to be the better fit for most software engineering and everyday knowledge work. Carbon, by contrast, is aimed at tasks that stress a model's intelligence, such as complex multi-step coding and open-ended scientific research.

In practice, that gives Google two premium models with distinct jobs: Argon for the bulk of routine work, and Carbon for the hardest problems customers bring to it.

One prompt, a team of agents

Alongside the release, Google is previewing a swarm feature on Antigravity aimed at long-running work with AI agents. As The Verge describes it, the feature lets a single prompt be turned into a coordinated team of agents — multiple AI instances working together on a task rather than one agent handling it step by step.

The pairing is deliberate. Carbon is built for long-horizon work, and swarms of agents are one of the more demanding ways to use a model, since coordinating many agents over an extended task compounds the need for reliable reasoning.

Aggressive introductory pricing

Google is pricing Carbon at $2 per million input tokens and $10 per million output tokens during an introductory period. According to The Verge, that is roughly a fifth of the cost of OpenAI's GPT-6 Astra and a quarter of the cost of Anthropic's Fable 5.2, which launched the day before Carbon's rollout.

Those prices are set to double at the end of November when the introductory window closes, but Carbon will still be significantly cheaper than its frontier rivals on a per-token basis. The Verge adds an important caveat: per-token prices do not translate directly into real-world costs, because different models can consume very different numbers of tokens to finish the same job.

Benchmark position

Benchmarks that Google cited when it first announced Carbon suggest the model is among the most powerful on the market across most metrics, though it trails Anthropic's Fable 5.2. The Verge notes the standard caution here: vendors routinely select the benchmark results that cast their own models in the best light.

Why it matters

Carbon's release, landing so soon after Argon, marks what The Verge calls a decisive return for Google to the very front of AI research. For months, frontier development had looked like a two-horse race between Anthropic and OpenAI; two major Google launches in consecutive months change that picture and restore a third serious contender at the top.

The pricing matters too. By undercutting rivals sharply on a per-token basis — even after the promotional period ends — Google is putting pressure on competitors' margins and could trigger broader price competition, making frontier-class models cheaper for developers building on top of them.

Finally, the swarm feature signals where Google believes agent development is heading: away from single agents solving problems sequentially and toward fleets of coordinated agents tackling large, long-running tasks. If that approach catches on, the platforms that make managing agent teams straightforward — which is what Google is attempting with Antigravity — will be well positioned to capture the resulting developer activity.