Google Opens Gemini 4 Carbon and Its Agent Swarm Preview

By Priya Raman

Google is rolling out Gemini 4 Carbon at $2 per million input tokens and $10 per million output tokens, with those introductory rates set to double at the end of November. The opening price is the release’s hardest number, but it does not settle the cost of a finished task.

Google prices Carbon at roughly one-fifth of OpenAI’s GPT-6 Astra and one-quarter of Anthropic’s Fable 5.2, which launched the day before Carbon. Even after November’s doubling, Google says Carbon remains significantly cheaper per token. Token consumption can differ across models for the same work, however, so a lower unit price may not lower the bill.

Who can use Carbon today?

In a launch blog post, Google said Carbon becomes available through the Gemini API in Google AI Studio and Gemini for business, covering all paying developers and enterprises. Google AI Ultra subscribers can also use it in the Gemini app, AI Mode in Search, and on the Antigravity development platform.

Google did not say when, or whether, lower-tier subscribers and free users will gain access, which makes general public a broader description than the published availability list. For anyone outside those named groups, the rollout has no stated endpoint.

How does the swarm change the workflow?

Antigravity is also getting a preview of a swarm feature. A user can submit one prompt and have Google convert it into a coordinated team of agents. That changes the workflow from one prompt addressed to one agent toward one prompt addressed to a coordinated team for long-horizon work.

Google did not disclose how many agents the team can contain, how they exchange work, what a person can inspect, or how the system measures completion. Nor did it provide a separate schedule for the swarm beyond describing it as a preview. The practical interface claim stays narrow: one prompt enters, and a coordinated team emerges.

What is Carbon meant to do?

Google said Carbon “pushes the frontier even further” than Argon, its other frontier model, which launched only the previous month. The company also calls Carbon its strongest model yet and says it can compete with Anthropic and OpenAI’s best systems.

That supports Google’s claim of a decisive return to the frontier after months when development looked like a two-horse race between Anthropic and OpenAI. The company is presenting two models for two levels of demand, rather than replacing its broad-purpose option outright.

Google expects Argon to remain the ideal choice for the majority of software engineering and knowledge work. It describes Carbon as “built for tasks that test the limits of a model’s intelligence,” including complex, multi-step coding and open-ended scientific research. The division separates general-purpose work from harder assignments with long chains of dependent steps.

Carbon is already familiar to a narrow test audience: Google released it to a select group of cybersecurity defenders two weeks before the general rollout. The announcement does not report what those defenders built, how their results compared, or whether the earlier group supplied evidence for the public benchmark claims.

What do the benchmarks show?

Google says the benchmarks cited at Carbon’s first announcement place it among the market’s most powerful models across most metrics, behind Fable 5.2.

The announcement supplies no scores, test sets, judging method, or independent verification. That omission matters because model companies can and do select benchmark runs that favor their tools. One of the most powerful remains a vendor claim, not a reproducible verdict.

The release sequence also compresses Google’s attempt to establish both a default model and a frontier option. Carbon followed Argon by one month and reached defenders two weeks before public availability. For developers choosing where to start, Google’s own guidance still points to Argon for most software engineering and knowledge work, while Carbon targets the harder end.

Is the lower price enough?

Per-token economics favor Carbon now, but the comparison has a trap. If one model uses more tokens to finish the same task, its lower unit price can still produce a higher total; Google itself cautions that the price advantage does not necessarily translate into lower task costs. Workload-level tests, absent from the announcement, would provide the relevant comparison.

For now, Google is running two tracks: Argon for most software engineering and knowledge work, and Carbon for complex coding and scientific research. The end-of-November rate increase and any expansion of the Antigravity swarm preview will test whether Carbon’s cheaper tokens and coordinated agents deliver measurable task-level value, rather than merely a stronger launch claim.