AI & DEV NEWS · · 4 min read

Gemini 4 Carbon is GA, and the $2 price has an expiry date

Google made Gemini 4 Carbon generally available on 10 November, with a /swarm multi-agent preview in Antigravity. The intro price is $2 in and $10 out per million tokens, and it won't last the month.

What shipped on 10 November, and where Argon still wins

Per Google's announcement, Carbon is on the API in Google AI Studio and Gemini for business for all paying developers and enterprises. It is rolling out to AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search. Google also says defenders got it two weeks earlier through its Fairwind Program, its cybersecurity access scheme.

The part we like is the candour. Google says Gemini 4 Argon stays the right pick for most software engineering and knowledge work. Carbon is for tasks that stretch a model, and it spends more tokens than Argon to explore, reason and run verification loops. A vendor telling you not to use its flagship for everything is rare.

SourceCodester's roundup adds that GitHub made Carbon generally available in Copilot the same day, for Pro+, Max, Business and Enterprise, with a gradual rollout. If your team lives in Copilot, you can probably try it without touching an API key. Per-model billing there has its own quirks, which we covered in our Copilot Fusion piece.

The #1 rankings are Google's claim, and the cost framing is narrow

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model. Those benchmarks come from third parties, but the results are reported by Google. As we read it, "external" excludes Google's own models, so the claim doesn't tell you whether Carbon beats Argon per dollar.

On a new Multimodal Knowledge Agent benchmark built with Artificial Analysis, Google puts Carbon at #2, with average cost per task more than 6X lower than the leader. That's an honest second place.

One roundup also says Carbon leads Terminal-Bench 4.0, and that Claude Fable 5.2 keeps FrontierCode, FrontierSWE and scientific research. We haven't verified that. If it holds, Carbon leads on automation-style work and doesn't sweep the board.

The $2 and $10 price is introductory, and per token isn't per task

Google lists Carbon at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens. The roundup reports that this runs through 30 November and rises to $4 and $20 in December. It also says Argon doubles on 1 December. The same roundup lists Fable 5.2 at $8 and $40, which is where its "four times cheaper" headline comes from. After the rise, the list-price gap is 2x.

Budget at $4 and $20, not at the launch rate. And remember that Carbon burns more tokens than Argon. At the same rate, a hard task can still cost more on Carbon. In the part of Google's post we could read, we found no Carbon-versus-Argon cost per task. You'll have to measure that yourself.

/swarm is a preview, and the demos are hand-picked

Google describes /swarm in Antigravity as one prompt becoming a coordinated team of agents you can watch in real time. Carbon takes the roles that need deep reasoning: architecture, coordination and auditing. Faster, lighter models work in parallel isolated workspaces, and adversarial verifiers test every change.

The showcase is big. One run turned seven planning PDFs into a 906K-triangle Three.js reconstruction of an office, across 10 overnight loops. It's impressive, and it's Google's pick of its best run. "Overnight" also hints at the token bill, and we didn't see a price for a swarm run.

The pattern is a planner, cheap workers and a checker. It's useful, but it hides control flow from you. If you need runs you can inspect and resume, our LangGraph teardown covers the explicit alternative. Treat /swarm as a look at where IDEs are heading, not something to build on yet.

What we'd do before 30 November

Nothing here needs a migration. It needs a small, honest test. If your automations call a model from n8n or Make, swapping it is usually a dropdown, so the test is cheap. Our AI workflow automation page covers how we wire those up.

  • Pick a few dozen real tasks from your own repos or workflows, including ones that failed before.
  • Run Argon and Carbon on each. Record pass rate, tokens and dollars at the $4 and $20 rate.
  • Route only the tasks where Carbon wins by a clear margin. Leave Argon as the default.
  • Try /swarm on a throwaway repo, and don't point it at production.
  • Put the December rate in your budget now.

Sources

  1. Gemini 4 Carbon is now available · Google
  2. AI Model Releases & News, 11 November 2026: Top Stories · AIToolsRecap
  3. .NET 11 Released, Critical Vite Dev Server Bug, Gemini 4 Carbon: Developer News for November 11, 2026 · SourceCodester

Want this working in your stack, not just in a demo?

We are two senior engineers who build ai workflow automation with a fixed price per written scope.

Keep reading.