Gemini 4 Carbon: What It Is, What It Can Do, Who Can Use It and How Much It Costs
On November 10, 2026, Google opened access to Gemini 4 Carbon — in the company’s own words, its strongest model. Here’s what it is, how it differs from regular Gemini, who can use it and how much it costs.
In short: Carbon is the top model of the Gemini 4 family, built for the hardest tasks: large coding projects, teams of AI agents, cybersecurity and science. In the Gemini app and AI Mode in Search it’s available only to Google AI Ultra subscribers for now; developers get it through the paid API. Google hasn’t announced free access.
What Gemini 4 Carbon is
The Gemini 4 family has two main models:
- Argon — according to Google, the ideal choice for most software engineering and knowledge work;
- Carbon — for tasks that test the limits: it thinks longer and uses more tokens than Argon, but handles long, multi-step tasks better.
Carbon first went to cybersecurity defenders through the invite-only Fairwind Program in late October, and on November 10 it opened to subscribers and developers.
What it can do: the /swarm of AI agents
The headline feature is /swarm (in preview) in Google’s Antigravity development environment. A single prompt turns into a team of agents: Carbon plans the architecture, coordinates the work and checks the result, while faster, lighter models build their parts in parallel in isolated workspaces. You can follow the progress in real time.
In its demos Google showed Carbon with such a swarm building an interactive 3D exhibition about the Eiffel Tower, reconstructing a 3D model of an office building with almost a million triangles (906,000) over 10 overnight loops, making a small open-world game, and modelling a 1780 fortepiano.
Google also highlights:
- visual understanding — dense documents, charts, blueprints and tables;
- cybersecurity — according to Google, Carbon leads SEC-bench Pro, ExploitGym and CWE-bench; Google’s partner Wiz used it to find previously unknown vulnerabilities at two SaaS providers in under two hours;
- honesty — on a 375-task agentic honesty benchmark, Carbon passed 92% of tasks versus 70% for Argon.
How strong it is
According to Google, Carbon ranks #1 on CursorBench 4.0 (coding), Perplexity’s WANDR and Zapier’s AutomationBench — at more than 2x lower cost per task than the next best model from another company. Google gives no specific scores, and all these results come from Google and its partners; there are no independent tests yet.
On safety, Google reports that in adaptive red-team testing Carbon blocked 99% of attack trajectories, and the per-attempt breach rate fell to 0.8% (3.1% for Argon).
How to try it and what it costs
| Where | Who can use it |
|---|---|
| Gemini app, AI Mode in Search | Google AI Ultra subscribers |
| Google Antigravity, Google AI Studio | AI Ultra subscribers (to try it) |
| Gemini API (AI Studio, Android Studio) | developers on the paid tier |
| Gemini for Business | companies |
API price: until November 30, 2026 the introductory price is $2 per million input tokens and $10 per million output tokens. From December 1 it’s $4 and $20 respectively.
For scientists, Google offers six months of free access through Antigravity — by application on Google’s site, subject to eligibility.
No subscription?
For everyday tasks — writing, email, translation, simple code — smaller models, including free ones, are enough. Carbon is worth it when the task is big and multi-step.
And if you want AI without subscriptions or an internet connection, you can run it on your own computer with Ollama: open models don’t reach Carbon’s level, but they handle everyday tasks fine.