gemini 4 carbon: Google Opens Its Top Model at $2/$10 With Agent Swarms

The gemini 4 carbon model is finally out for everyone who pays: Google opened its heaviest model to developers, enterprises and AI Ultra subscribers on November 10, two weeks after a limited release to vetted cyber defenders. It ships at an introductory $2 per million input tokens and $10 per million output tokens, plus a preview of a "/swarm" mode that turns one prompt into a team of agents.

What actually happened

Google first showed Carbon in late October and gave it to a small group of security defenders. On Tuesday it went to general availability for paying developers and enterprises, according to The Verge and The Decoder. Google calls it its strongest model and pitches it at "tasks that test the limits of a model's intelligence," like long multi-step coding, debugging and open-ended research.

The positioning is unusual. Google says Argon, the frontier model it launched last month and which we covered in gemini 4 argon, stays "the ideal choice for the majority of software engineering and knowledge work." The Decoder points out that both models share the same token price, but Carbon spends more tokens thinking and running verification loops, so the real bill per task is likely higher.

Pricing is the flex. The introductory rate is $2 input and $10 output per million tokens, and it doubles to $4/$20 when the intro period ends at the close of November. The Verge estimates even the launch price is around a fifth of OpenAI's GPT-6 Astra and a quarter of Anthropic's Fable 5.2 per token, while warning that per-token price is not the same as cost per task.

The /swarm preview

The new toy is "/swarm" in Antigravity, Google's agentic coding platform. You give one prompt and it spins up a coordinated team of agents you can watch in real time. In Google's description, Carbon takes the deep-reasoning roles like architecture, coordination and auditing, while lighter models work in parallel sandboxes and act as adversarial checkers.

Google's demos include a 3D game where you ride a GBike around the Mountain View campus as a pelican, a reconstruction of the Eiffel Tower's construction from historical photos, and a playable 18th-century fortepiano. Fun, but The Decoder notes the industry is still arguing about whether agent swarms are worth their cost beyond raw speed.

Benchmarks, with a grain of salt

Google says Carbon ranks first on CursorBench 4.0 with a score just under 60%, ahead of Fable 5.2 and Argon, and also leads Perplexity's WANDR and Zapier's AutomationBench at less than half the per-task cost of Fable 5.2. Those are vendor-picked results.

Independent numbers are closer. In Artificial Analysis testing cited by The Decoder, Carbon cost $7.76 per solved task at standard pricing, and Fable 5.2 at its High setting scored 62 on the overall index versus 60 for Carbon. So gemini 4 carbon is a top-tier option, not a runaway winner.

What Polymarket thinks

The best AI model market resolves on the arena.ai Text leaderboard (style control off) at noon ET on November 30. At about 06:01 UTC on November 11, Google traded near 81.5% (bid 0.80 / ask 0.83), Anthropic near 15.5% and OpenAI around 1%. Google gained roughly 6 points in a day.

That market is about leaderboard votes, not price or benchmarks, so it can disagree with Artificial Analysis. We walked through the setup in november ai model odds before this launch.

What to do as a reader

If you build with LLMs, test gemini 4 carbon on your own hardest tasks during the cheap intro window and track total cost per finished job, not just token price. Compare it with Argon on the same prompts, because Google itself says Argon wins on most routine work. Check the official Gemini API docs for current limits and pricing.

If you trade the AI-model markets, remember that arena rankings move with user votes and new releases, and a 15-point underdog can flip the board with one strong launch. Nothing here is investment advice.