Google's Gemini 4 Carbon Is Now Generally Available — and It's a Quarter of Claude Fable 5.2's Price

Two weeks ago, Google unveiled Gemini 4 Carbon — its most powerful model — and then told most developers they couldn't have it. Carbon was reserved for a subset of cybersecurity defenders through Google's Fairwind Program while safeguards were still being tested. Today, November 10, 2026, that changes: Carbon is generally available. If you are a paying developer, an enterprise customer, or a Google AI Ultra subscriber, you can use it now.

I have not run Carbon hands-on yet — nobody outside the Fairwind group had access before today — and this article is built from Google's announcement and the VentureBeat and 9to5Google coverage of it, checked November 10, 2026. Where a claim is Google's, I'll say so, because a launch-day benchmark is a vendor claim, not an independent result.

What actually changed today

The model itself is the one Google announced on October 28: a frontier model built for long-running agentic workflows, software engineering, scientific reasoning, and cybersecurity. What changed is the door. On October 28, Carbon was a press release with a velvet rope in front of it. Today it is a model you can call.

It is available through the Gemini API (including AI Studio and Android Studio), Gemini for Business, the Gemini app, Search AI Mode, and Antigravity — for Google AI Ultra subscribers. That breadth matters: Google is not drip-feeding Carbon through one channel. It is putting its strongest model in front of paying developers and enterprise users on day one of general availability.

The benchmarks Google is leading with

Google says Carbon ranks #1 on CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. Those are the benchmarks that map most directly onto real developer work: agentic coding, research workflows, and automation. Google also claims frontier performance on long-horizon agentic tasks and scientific reasoning.

Read those numbers the way you should read any launch-day claim: they are Google's, on benchmarks Google chose to highlight, and nobody independent has reproduced them yet. But the choice of benchmarks is itself informative. Google is not leading with a general knowledge test. It is leading with the tests that measure whether a model can do sustained, multi-step work without falling apart — which is exactly the job Carbon was built for.

The price is the real story

Here is the number that will move workloads. Carbon costs $2 per million input tokens and $10 per million output tokens as an introductory price until November 30, 2026 — then $4 input and $20 output. Cached input gets a 95% discount.

Now compare that with what landed yesterday. Anthropic's Claude Fable 5.2, launched November 9, costs $8 per million input tokens and $40 per million output tokens. At introductory pricing, Carbon is a quarter of Fable's price. Even at Carbon's regular $4/$20 pricing after November 30, it is half of Fable's.

And the performance gap, on the one independent scoreboard we have, is small. Artificial Analysis scores Fable 5.2 at 65 and Carbon at 60. Five points, for a quarter of the price at launch pricing. That is not a rounding error in a budget — it is the difference between a model you reserve for your hardest task and a model you can afford to point at everything.

There is a catch in the calendar, and it is worth stating plainly: the $2/$10 price expires November 30. If you are evaluating Carbon, evaluate it at $4/$20 too, because that is the price you will actually be paying in December. Even then, the math holds up well against Fable — but don't build a budget on a price that has a three-week expiry date.

/swarm: one prompt, a team of agents

Alongside general availability, Google is previewing /swarm in Antigravity: type one prompt and a coordinated team of agents works on it together, with their progress tracked in real time.

Multi-agent orchestration is the feature every coding assistant is racing toward right now, and most implementations still feel like a demo with a nice UI. The honest question for /swarm is the same one I would ask of any of them: does the team actually coordinate, or does it just run several agents in parallel and hand you the merge conflicts? That is a hands-on question, and I can't answer it from an announcement. But if you are already in Antigravity, it is the first thing worth trying — because a swarm that works changes what a single prompt can do, and a swarm that doesn't is just more tokens.

Carbon vs Fable 5.2: how to choose

Yesterday's launch and today's are now a genuine choice, and the choice is simpler than the benchmark tables make it look:

• Pick Carbon if price shapes your usage. At $2/$10 until November 30 and $4/$20 after, with a 95% cached-input discount, it is the cheaper way to run long agentic workflows — and long workflows are where the token bill actually hurts.
• Pick Fable 5.2 if you need the top score and the Anthropic ecosystem. Artificial Analysis has it five points ahead, and if your workflow is already built around Claude Code, switching costs are real.
• Try both on the same hard task. This is the honest test, and it costs almost nothing this week: give each model the hardest task on your plate — the multi-file refactor or the long debugging session that usually needs two or three attempts — and see which one finishes in one.

What I would not do is pick based on the launch headlines alone. Carbon at #1 on three benchmarks and Fable at 65 on Artificial Analysis are both vendor-adjacent numbers. The models are close enough that your own task is a better judge than either scoreboard.

What it means for developers

The practical takeaway is short. Thirteen days ago, Carbon was a model you could read about but not use. Today it is a model you can call from the Gemini API, from AI Studio, from Android Studio, and from Antigravity — at a price that undercuts the model that launched yesterday by a wide margin.

If you have been waiting to try Carbon, the wait is over. If you have been budgeting around Fable 5.2, run the same task through Carbon first and compare the bill. And if you use Antigravity, try /swarm on something real — a coordinated agent team is either the next step in how we build software, or an expensive way to watch several agents do the same work. The only way to know which is to try it.

Would you switch your hardest coding task to Carbon this week, or does Fable 5.2's five-point lead keep you on Anthropic?

Sources: Google announcement via blog.google, November 10, 2026; VentureBeat coverage, November 10, 2026; 9to5Google coverage, November 10, 2026; Anthropic Fable 5.2 pricing and Artificial Analysis scores as reported November 9, 2026. All capability, benchmark and pricing claims are Google's or Anthropic's, stated as such. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.