Google DeepMind opened Gemini 4 Carbon to everyone on November 10. Google positions it as its strongest model so far, and it's aimed at the hardest work: long-running AI agent jobs that need repeated checking, plus cybersecurity. Developers can call it through the API, and individual users need a Google AI Ultra subscription to get at it.
What happened
Two weeks ago, Carbon was only available to a group of security teams in Google's Fairwind program. Now it's open across the board:
- Developers and businesses: You can pick it on the paid API tier, whether you work in Google AI Studio or through Gemini for business, and it's also available in Android Studio.
- Individual users: Google AI Ultra subscribers can use it in the Gemini app, in AI Mode in Search, and in Google Antigravity.
- Security teams: You can apply to Fairwind to get Carbon for cyber, a version built for vulnerability research.
- Researchers: The program that gave scientists free access to Argon now covers Carbon too, with rolling reviews.
A preview of /swarm in Antigravity launched alongside it. One sentence starts a team of AI agents that split up the work. Carbon takes the demanding roles like architecture, coordination and review, while lighter models write code in their own isolated workspaces at the same time. A dedicated checker agent reviews every change.
Carbon vs. Argon
Google didn't use Carbon to replace Gemini 4 Argon, which came out not long ago. It split the work between the two.
| Gemini 4 Argon | Gemini 4 Carbon | |
|---|---|---|
| Best for | Most coding and everyday knowledge work | Multi-step complex coding and debugging, open-ended research, vulnerability research |
| How it works | Normal reasoning | Spends more tokens exploring and double-checking |
| API price (per million tokens) | Input $2 / output $10 | Same price as Argon until November 30, then input $4 / output $20 |
Use Argon by default. When a job is too big for it and needs to run on its own for a long stretch, switch to Carbon. Carbon thinks for longer and uses more tokens per task, so keep an eye on your bill.
What it does well
Google's results focus on tests that resemble real workflows:
- AI agent work: It ranks first on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, and each task costs less than half of what the second-place outside model spends.
- Reading complex files: On the multimodal knowledge agent test (MKA) that Google and Artificial Analysis just launched, it ranks second, with an average cost per task more than six times lower. The test checks whether an agent can make sense of large files full of charts, drawings and tables.
- Cybersecurity: It leads in all three categories: finding vulnerabilities, exploiting them and patching them. With its help, a Google security team showed that two flaws in a widely used open-source package were genuinely exploitable, which helped maintainers raise the severity and ship fixes faster.
- Honesty: It passed 92% of Google's 375-task agent honesty test, compared with 70% for Argon. When data is missing, it usually says so instead of inventing an answer.
These are Google's own numbers, and there hasn't been much independent testing yet.
What this means if you use AI agents
If you run AI agents on big projects, like migrating a whole repo, tracking down bugs across modules, or turning a pile of PDFs into a report, Carbon is worth a try. Its pitch is that it can grind through a long chain of steps and then check its own work.
The easiest way to test it is to have your AI agent grab a paid API key from Google AI Studio, switch the model to Gemini 4 Carbon, and run it on a long task that Argon or your current model never finishes. Compare how far it gets and what it costs before you commit.
Watch out
- The price doubles after November 30. The current pricing is a limited-time offer, so budget with $4 / $20 for anything long-term.
- The personal bar is high. Only Google AI Ultra subscribers can use it in the Gemini app. The regular paid tiers can't.
- No free API tier. It's only for paying developers.
- Security features are restricted. Full vulnerability research requires applying through Fairwind, and the public version blocks attack-style requests more strictly.
Comments (0)
No comments yet. Start the conversation.