Gemini 4 Carbon: What It Does, /swarm and the Price Doubling

Google released Gemini 4 Carbon on November 10, 2026, roughly a month after Gemini 4 Argon, and the pitch is narrower than the model name suggests. Carbon is not positioned as a replacement for Argon. Google says Argon “remains the ideal choice for the majority of software engineering and knowledge work”, while Carbon is built for tasks that stress a model hardest: long multi-step coding and debugging, open-ended scientific work, and cybersecurity.

If you are weighing it against the assistants you already pay for, our comparison of ChatGPT, Gemini and Claude is the wider context; this piece sticks to what Google announced on November 10.

Where you can actually use it

Per Google’s developer post, Carbon is available through the paid tier of the Gemini API in Google AI Studio and Android Studio, and in Gemini for business for paying developers and enterprises. Google’s launch post adds that it is rolling out to Google AI Ultra subscribers in the Gemini app, in AI Mode in Search, and in Antigravity.

What Google did not say is when, or whether, Carbon reaches lower-paid tiers or free users. The Verge noted that it is unclear, so treat “free” claims as unconfirmed.

The pricing, which has a deadline attached

Carbon launched at the same introductory rate as Argon: $2 per million input tokens and $10 per million output tokens. The Verge reports that the introductory period ends at the end of November, after which those prices double to $4 and $20 per million tokens. Even at the higher rate, that is well below the frontier models it competes with on a per-token basis, though per-token pricing is a poor proxy for cost per task: two models can finish the same job using very different numbers of tokens.

If you are evaluating Carbon for a production workload, run your own cost test before the introductory rate expires rather than after.

What /swarm changes

Google is previewing /swarm inside Antigravity, its agent-first development environment. The Antigravity team’s explanation is more specific than the marketing: typing the command hands your prompt to a coordinator that spins up specialised parallel workers in isolated git worktrees so edits do not collide, plus adversarial verifiers whose job is to run tests and challenge the resulting diffs. Agents coordinate on a shared, append-only message board and park at zero token cost when idle.

Two practical consequences. First, agents can be pinned to different models, so you might put Carbon on coordination and auditing while cheaper, faster models make routine file edits. Second, the tool ships with a live board, a member roster and a graph replay of how the team got from prompt to code, which is the difference between a multi-agent system you can debug and one you have to trust.

About those benchmark numbers

Google says Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, and claims more than 2X lower cost per task than the next best external model. Those are Google’s characterisations of its own benchmarks. The Verge points out the obvious caveat: vendors choose benchmarks that flatter their models, and its own summary of Carbon’s earlier benchmark disclosure placed the model behind Anthropic’s Fable 5.2 on most metrics. We have not independently tested Carbon and are not repeating the rankings as fact.

Who should care right now

Teams already on Argon: little reason to switch by default. Google itself says Argon remains right for most work.
Anyone with long-running agent workflows: Carbon’s intended use case is exactly the messy, multi-file task where a cheaper model gets lost, and /swarm is the more interesting of the two announcements.
Budget owners: the calendar matters more than the leaderboard. Model choice at $10 versus $20 per million output tokens is a finance decision as much as an engineering one.