Gemini 4 Carbon: how to use it, pricing and /swarm
2026-11-10 · 4 min read
You can use Gemini 4 Carbon today in two ways: through the paid Gemini API in Google AI Studio, or with a Google AI Ultra subscription in the Gemini app, AI Mode in Search and Antigravity. Google's posts don't mention a free tier, and the API price doubles after November 30.
Google opened the model up this morning. @Google posted that it's now available and pitched it against its sibling: "Carbon is the model you bring in for your hardest problems." Two weeks ago, when we wrote what Gemini 4 Carbon is and who can use it, only vetted security teams had access and Google had given no date.
Where to find Gemini 4 Carbon
Google's announcement and its developer post list these routes:
| You are | Where you get Carbon | What you need |
|---|---|---|
| A developer | Gemini API in Google AI Studio and Android Studio | Paid API tier |
| A consumer | Gemini app, AI Mode in Search | Google AI Ultra |
| A coder using agents | Google Antigravity | Google AI Ultra |
| A business | Gemini for business, by API or the model drop-down | Paid plan |
| A security team | Fairwind Program | Approved application |
Google says the Ultra rollout is still in progress, so the model may take a little while to show up in your account. Scientists can also apply for six free months of Antigravity's top tier, though Google picks who gets it.
How to try it
In the Gemini app, open the model picker. 9to5Google reports that 4 Argon is labeled "Code & analysis" and 4 Carbon is labeled "Most complex work." Carbon is an Ultra feature for now. AI Pro subscribers keep Argon, which 9to5Google says stays available alongside it.
For the API, you need a paid account in Google AI Studio. The AI Studio link in Google's post opens a model named gemini-4-carbon, so that's the string to look for. If your code already calls Argon, I'd expect swapping the model name to be most of the work.
Google hasn't published a context window, output limit or usage caps in either post, so check the model card before you plan a big job around it.
Gemini 4 Carbon pricing
| Per million tokens | Until November 30, 2026 | After |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
That's the same introductory rate Argon launched with. For comparison, Anthropic priced Claude Fable 5.2 at $8 input and $40 output yesterday, so Carbon comes in at half of that even at its full price.
That leaves about three weeks at half price, enough time to run your hardest prompts through both models and compare the bills. Budget for $4 and $20 anyway, because that's what you'll pay from December on.
What /swarm does in Antigravity
Antigravity also gets a preview command called /swarm. Google describes it as "turning a single prompt into a coordinated team of agents you can track in real time."
It splits the work by cost. Carbon handles the design, the coordination and the final audit. Faster, cheaper models do the bulk of the building in parallel, each in its own isolated workspace, and separate verifier agents try to find holes in the result.
Google hasn't said how the agents share context, what happens when two of them change the same file, or what a swarm run costs in quota. Until it does, I'd try it on a side project before pointing it at anything that matters. We compared the two setups in agent teams vs a single agent.
Is it worth switching from Argon?
Usually not, and that's Google's own position. Its post says Argon "remains the ideal choice for the majority of software engineering and knowledge work," and Carbon is for tasks at the limit of what a model can do.
Most of the numbers in the post are about agents. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at less than half the cost per task of the next best outside model. It also passed 92% of 375 "agentic honesty" tasks, against 70% for Argon. Those tasks check things like whether the model refuses to invent missing data instead of quietly filling it in.
All of that is Google grading its own model. The honesty figure is the one I'd watch, because an agent that admits it's stuck is worth more to a small team than one that scores two points higher and makes things up. This morning's post on AlphaProof Code shows what Carbon looks like when it's pointed at a hard, checkable job.
What this means for businesses adopting AI
For everyday work like drafting emails or answering customers, Argon or a cheaper model is still the right call. Carbon is worth a look if you have one process where mistakes are expensive and the steps run long. If you're not sure which of your processes fit that description, New Face Design offers a free process audit.