# Google prices Gemini 4 Carbon at a quarter of Anthropic's Fable 5.2 for now. It still ranks 5 points behind

Google's new top model costs $2 in and $10 out per million tokens until Nov. 30. Anthropic's Fable 5.2 leads the main index, but Carbon wins some cost-per-task tests.

By Patrick Karsh · 567 Business
Published: 2026-11-10T21:52:04.806994Z
Updated: 2026-11-10T21:52:14.763449Z
Canonical: https://567business.com/news/gemini-4-carbon-price-war-fable

Cite citation_url. Preserve source links, dates and qualifications. date_modified is an edit date, not proof that every claim was rechecked. Third-party images and quoted material retain their own rights. Resolve relative links against https://567business.com/news/gemini-4-carbon-price-war-fable.

Google released Gemini 4 Carbon to the public on Tuesday, and its sales pitch is price. Paying developers can use its most capable model for $2 per million input tokens and $10 per million output tokens through Nov. 30, [Google said](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/). That is a quarter of what Anthropic charges for Claude Fable 5.2, the model it shipped one day earlier, [according to VentureBeat](https://venturebeat.com/ai/google-gemini-4-carbon-is-now-publicly-available-and-it-undercuts-claude-fable-5-2-and-opus-5-5-on-api-price).

The catch is the ranking. On the most-watched independent leaderboard, Fable 5.2 sits first and Carbon sits five points behind. So the question for buyers is whether the cheaper model is close enough, and early data points in both directions.

## What Google released

Carbon is Google's strongest model, built for long agentic jobs such as multi-step coding, debugging and research, wrote Koray Kavukcuoglu, Google's chief AI architect. Google first gave it to a small group of cyber defenders through its "Fairwind" program two weeks ago. As of Tuesday it is open to all paying users of the Gemini API and rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and the Antigravity coding tool.

Google is also previewing a feature called `/swarm` in Antigravity. It turns one prompt into a team of agents, with Carbon handling planning and review while lighter models do parallel work. [The Verge](https://www.theverge.com/ai-artificial-intelligence/1030787/google-gemini-4-carbon-public-rollout) called it an agent "swarm" feature and the headline addition of the rollout.

The low price does not last. Google's own footnote says the rate rises to $4 and $20 per million tokens after the introductory period ends on Nov. 30. At that point Carbon costs half of Fable 5.2, not a quarter.

## The price gap

```chart
type: bar
title: Output price per million tokens, frontier models
note: Standard API output rates. Carbon's $10 rate is introductory and ends Nov. 30; it then rises to $20.
prefix: $
source: [VentureBeat](https://venturebeat.com/ai/google-gemini-4-carbon-is-now-publicly-available-and-it-undercuts-claude-fable-5-2-and-opus-5-5-on-api-price); [Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/)
---
Model | Output price
OpenAI GPT-6 Astra | 50
Anthropic Claude Fable 5.2 | 40
Anthropic Claude Opus 5.5 | 20
Gemini 4 Carbon (after Nov. 30) | 20
Gemini 4 Carbon (introductory) | 10
```

Input prices follow the same pattern: $10 for GPT-6 Astra, $8 for Fable 5.2, $4 for Opus 5.5 and $2 for Carbon at launch.

## The ranking that still favors Anthropic

Artificial Analysis, an independent group that runs a widely cited composite index, put Fable 5.2 in first place on Monday. In its words, Fable "at max effort scores 65, 5 points ahead of Gemini 4 Carbon (high, 60) and 6 ahead of Claude Opus 5.5 (max, 59)." The post drew about 2,800 likes and 246,000 views.

https://x.com/ArtificialAnlys/status/2119878995386982508

The Verge also reported that benchmarks place Carbon behind Fable 5.2. Some users on X were blunter. A post from the account Angaisb_, liked about 750 times, said Fable 5.2 "completely destroyed" Carbon.

| Model | Artificial Analysis index | Output price ($/1M) |
| --- | --- | --- |
| Claude Fable 5.2 (max) | 65 | 40 |
| Gemini 4 Carbon (high) | 60 | 10 (20 after Nov. 30) |
| Claude Opus 5.5 (max) | 59 | 20 |

Sources: [Artificial Analysis on X](https://x.com/ArtificialAnlys/status/2119878995386982508); [VentureBeat](https://venturebeat.com/ai/google-gemini-4-carbon-is-now-publicly-available-and-it-undercuts-claude-fable-5-2-and-opus-5-5-on-api-price).

## Where Carbon wins: cost per task

Google argues that the index misses the point for businesses. It says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench "at more than 2X lower cost per task compared to the next best external model." Those are tests built around real work, such as editing code inside an app or finishing tasks across business tools.

The X account TekonBurns posted the CursorBench numbers on Tuesday. By his account, Carbon at high effort scored 60.4% at $9.22 per task, while Fable 5.2 at max effort scored 60.0% at $32.79 per task. That is a near tie on accuracy at less than a third of the cost. We could not independently check the figures against Cursor's own page, but they match Google's claim of first place at more than double the cost efficiency.

https://x.com/TekonBurns/status/2120204822695805048

## The cost complaint

Not every user sees savings. The X user rohit3a posted on Tuesday that in his testing Carbon cost more than Fable 5.2 on high effort, despite lower token prices. He blamed the model for being "bad at hitting cache," which would mean it reads the same input again at full price. His post is one user's report, not a benchmark.

The Verge made the general point: price per token is not the same as price per task. Google itself says Carbon uses "more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops." A model that thinks longer can erase a low sticker price. VentureBeat put it this way: "The bigger question for buyers is how many tokens, retries and human corrections it takes to finish the job."

## Why the fight matters to the AI business

Coding agents are the most competitive part of this market, and Anthropic's Fable line is Google's main rival there. Anthropic is preparing for a public listing, and its rich valuation rests partly on its lead with developers, as our [Anthropic IPO story](/news/anthropic-ipo-valuation-selloff) explains. A Google model that is nearly as good at a fraction of the price is a direct threat to that pricing power.

Google can afford a price war in a way most labs cannot. It runs its own chips and data centers, and its parent has the largest pile of future AI spending commitments of any company, as our [look at $4.49 trillion in off-balance-sheet AI obligations](/news/ai-off-balance-sheet-4-trillion-bailey-bubble) shows. Selling access cheaply can make sense if it wins developers who will later pay more.

OpenAI, meanwhile, prices GPT-6 Astra above both. It is also spending its attention on safety after two lab incidents, covered in our [story on its new Control Room kill switch](/news/openai-control-room-35-seconds-hugging-face).

## What to watch before Dec. 1

The introductory price ends Nov. 30. After that, Carbon's list price doubles, and the gap with Fable 5.2 narrows from 4-to-1 to 2-to-1. Developers now have three weeks to test whether Carbon's lower token price survives its heavier thinking. The first wave of public cost-per-task tests, like Cursor's, suggests it can. The first wave of user complaints, like rohit3a's, suggests it depends on the job.