Google has made Gemini 4 Carbon, its most capable model so far, available to paying developers, businesses and subscribers. The company announced the release on Nov. 10. It sits above Gemini 4 Argon, which Google still recommends for most everyday coding and knowledge work.
Who can use it
Access depends on how you pay for Google’s AI:
- Developers and businesses can use it through the paid tier of the Gemini API, including Google AI Studio.
- Google AI Ultra subscribers are getting it in the Gemini app, AI Mode in Search and Google’s Antigravity coding tool, on a rolling basis.
- A cyber-focused version is limited to approved applicants in Google’s Fairwind Program, which gave early access to security defenders two weeks earlier.
- Scientists can apply for free six-month access, subject to eligibility, capacity and selection.
There is no mention of a free-tier option for Carbon. If you only use the free Gemini app, you should not expect to see it.
What it costs
Google lists an introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon. That price runs through Nov. 30, 2026. After that, it rises to $4 per million input tokens and $20 per million output tokens, double the introductory rate.
There is a wrinkle for budgeting. Google says Carbon uses more tokens than Argon on the same job, so the real cost per task can be higher than the per-token price suggests. Google also claims lower cost per task than rival models on some tests, but those are the company’s own comparisons.
What Google claims
The company says Carbon ranks first on several third-party coding and agent benchmarks, including CursorBench 4.0 and Zapier’s AutomationBench, and second on one multimodal agent test. It also says the model blocked 99% of simulated attack attempts from an automated red-team agent, with a per-attempt breach rate of 0.8%, down from 3.1% for Argon.
Treat those numbers with some care. Most of the results are reported by Google or its partners rather than independent testers, and Google itself says resistance to prompt injection, where hidden instructions trick an AI agent, “remains an area for continued improvement.” The protein-science result is also described as preliminary.
A new multi-agent feature
Google is also previewing a feature called /swarm in Antigravity. A single prompt becomes a coordinated team of AI agents, with Carbon handling design, coordination and auditing and faster models doing parallel work. The examples Google shows are its own demos, not independent evaluations.
What to do with this
For most people, nothing changes today. If you pay for Google AI Ultra, try Carbon on a hard task and compare it with the default model before relying on it. If you run an API budget, test both models on a sample of your real work and compare total cost, not just token prices, before the Nov. 30 price change.
More on Contoh
- Google’s Gemini 4 Argon Is Here. Who Gets It, What It Costs and What to Watch
- AI Coding Agents Fully Completed Only 26% of Real-World Tasks in a New Benchmark
- Revolut Cut Its AI Coding Costs About 90% by Hosting Its Own Model