Gemini 4 Carbon Pricing vs Claude Fable 5.2: The Real Math

Google made Gemini 4 Carbon generally available on November 10, 2026, at an introductory $2 per million input tokens — a quarter of Claude Fable 5.2's price. Output runs $10 per million.

Two weeks earlier, Carbon was a press release. Google had restricted it to a group of "cybersecurity defenders" through its Fairwind Program while safeguards were tested. Now any paying developer can call it.

What shipped, and where you can call it

Carbon is rolling out through the Gemini API in Google AI Studio and Gemini for Business for all paying developers and enterprises, according to Google's blog post as reported by The Verge. Google AI Ultra subscribers also get it in the Gemini app, AI Mode in Search, and on Antigravity. The dev.to breakdown by Paul Lovelace adds Android Studio to that list.

The Verge notes one open question: "It is not clear when — or if — Carbon will be available to lower tier subscribers or free users."

Vercel shipped it on AI Gateway the same day under the model name google/gemini-4-carbon, at 50% off through November 30, 2026. Vercel's changelog puts the context window at 1M tokens "for working with large codebases and long inputs," and states that AI Gateway "reflects provider pricing with no markup and does not charge a platform fee on inference, including on Bring Your Own Key (BYOK) requests."

If you want it inside an existing agent, Vercel's path is two commands:

npx vercel ai-gateway setup
# then select google/gemini-4-carbon in the agent

Vercel lists Claude Code, Codex and Cursor among the agents that route through it.

The pricing table, and the date on it

This is the part worth putting in a spreadsheet before you move any workload.

Model	Input / 1M tokens	Output / 1M tokens
Gemini 4 Carbon (intro, through Nov 30, 2026)	$2	$10
Gemini 4 Carbon (from Dec 1, 2026)	$4	$20
Claude Fable 5.2 (launched Nov 9, 2026)	$8	$40

Lovelace's write-up also reports a 95% discount on cached input for Carbon, which matters more than the headline rate for agent loops that resend the same system prompt and repository context on every turn.

The Verge frames the launch rate as "around a fifth the cost of OpenAI's GPT-6 Astra and a quarter the cost of Anthropic's Fable 5.2 model." It also adds the caveat that kills most naive price comparisons: cheaper per token "doesn't necessarily translate to lower costs as models can use different numbers of tokens for the same tasks."

So treat the table as a starting hypothesis, not a budget. The introductory price expires November 30. Anything you model today at $2/$10 you will be paying $4/$20 for in December — still half of Fable's rate, but a doubling you should see coming.

How close are they on quality?

Google says Carbon ranks #1 on CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench, per Lovelace's summary of the announcement. Those are vendor-selected benchmarks, and The Verge says so bluntly: "companies can and do cherry-pick benchmark results to show their tools in the best possible light."

The one third-party number in circulation comes from Lovelace, who cites Artificial Analysis scoring Fable 5.2 at 65 and Carbon at 60. The Verge's read of Google's own benchmark deck agrees on the direction — Carbon is "one of the most powerful on the market across most metrics, behind Fable 5.2."

Five points on one aggregate index, against a 4x price gap at launch rates. That is the whole decision, compressed.

Carbon is not Google's default recommendation

A detail that gets lost in the price story: Google is not positioning Carbon as the model you use for everything. The Verge quotes the company saying Argon — the frontier model Google launched only last month — "will likely remain the ideal choice for the majority of software engineering and knowledge work," while Carbon is "built for tasks that test the limits of a model's intelligence," such as complex multi-step coding and open-ended scientific research.

Read that as a routing instruction. If your workload is ordinary application code, Google's own guidance points at the cheaper, older model.

Swarm: one prompt, several agents

Alongside general availability, Google is previewing a "swarm" feature on Antigravity, its agentic development platform, which turns a single prompt "into a coordinated team of agents," per The Verge.

Lovelace poses the right test for it, and flags that he has not run it: does the team actually coordinate, or does it run several agents in parallel and hand you the merge conflicts? Nobody outside the Fairwind group had hands-on access before launch day, so every claim about swarm quality right now is a claim about a demo.

What to do this week

• Run one task twice. Take the multi-file refactor or long debugging session that usually needs two or three attempts, send it to Carbon and to your current model, and compare both the result and the token bill — not the per-token rate.
• Model December, not November. Build any switching case at $4/$20. The $2/$10 window closes on the 30th.
• Measure cache hits before you count on the 95% discount. If your agent rebuilds its prompt every turn, that discount does not apply to you.
• Check which model your routing actually needs. Google says Argon covers most engineering work; paying frontier rates for frontier-free tasks is the expensive mistake here.

The thing to watch next is whether independent evaluations hold the 5-point gap, and whether Carbon's December prices stay put once the introductory period ends.