P. 08  ·  Nov 13, 2026  ·  4 min read

Claude Fable 5.2 vs Gemini 4 Carbon on cost, quality and access

A well-kept tan leather suitcase, closed with two straps, with a small ink-blue stitched band on the handle.

Carbon is far cheaper per token today, but the bill for a finished task depends on the setting and on Carbon's December price rise.

Two flagship models, one day apart

Anthropic released Claude Fable 5.2 on November 9. Google made Gemini 4 Carbon widely available a day later.

Both are aimed at long jobs, where an AI plans, writes and fixes things on its own. Both makers say most teams should start with a cheaper model: Anthropic names Opus 5.5, and Google names Gemini 4 Argon. If a cheaper model falls short on your job, these two are the next step. The first question is always the same: which costs less?

ComparedClaude Fable 5.2Gemini 4 Carbon
Average cost per task (USD)$26.54max setting, per Intelligence Index task; measured$9.08high setting, per Intelligence Index task; measured$7.76high setting, per Intelligence Index task, December rate; measured$6.59medium setting, per Intelligence Index task, December rate; measured
Intelligence Index score65max setting; measured62high setting; measured60high setting; measured58medium setting; measured
Agentic coding (FrontierSWE v2, %)69.9%highest-available effort, safeguards on; vendor61.2%highest-available effort; vendor
Knowledge work score (MKA, %)57%Climb WSR, max setting; measured73.26%Overall WSR (every criterion scored); reported54%Climb WSR, high setting; measured71.89%Overall WSR (every criterion scored); vendor
Price per million output tokens (USD)$40standard rate, not fast mode; thinking tokens included; vendor$80fast mode, up to 2.5 times faster output; vendor$10introductory price until Nov 30; thinking tokens included; vendor$20from Dec 1; thinking tokens included; vendor
Price per million input tokens (USD)$8standard rate, not fast mode; vendor$16fast mode, up to 2.5 times faster output; vendor$2introductory price until Nov 30; uncached input; vendor$4from Dec 1, after the introductory price; uncached input; vendor
Time per task (minutes)47.8max setting, per Intelligence Index task; measured16.6high setting, per Intelligence Index task; measured25.5high setting, per Intelligence Index task; measured23.0medium setting, per Intelligence Index task; measured
Cloud platforms offering the modelAWS, Google Cloud, Microsoft Azurelaunch day; list not exhaustive; also Claude API; vendorGoogle AI Studio; Gemini for businessAPI access, paying developers and enterprises; vendor1 API providerreported

Data as of November 12, 2026. “Vendor” is what the maker states. “Measured” is an independent test. “Reported” is a third party’s figure that it did not measure itself.

What a finished task costs

On the price list, Carbon looks like the bargain. Fable costs $8 per million input tokens and $40 per million output tokens. A token is a small chunk of text, and models bill for what they read and write. Carbon's launch price is a quarter of that. It doubles on December 1, to half of Fable's price. I found no public price guarantee from either company, so I expect more changes.

A price per token is not a bill per task. The bill depends on how much a model writes while it thinks. Each model has a thinking setting: more thinking usually means better answers and a bigger bill. Artificial Analysis, an independent tester, ran both on the same jobs. At the high setting, an average task from its test set cost $9.08 on Fable and $7.76 on Carbon, using Carbon's December price. Use that to compare, not to budget. Averages hide big swings. Fable has built a simple cylinder for under $40, yet AI YouTuber Matthew Berman estimated $14,710 at pay-as-you-go prices for one game level. Failed tries add up: on Zapier's business-workflow test, Carbon at high is estimated at $2.12 per task, but $3.94 per task it gets right. I found no public rerun rate for Fable. The tester calls both models very verbose, and I suspect that long thinking eats most of Carbon's price edge.

“I would pick Fable at high for hard, long work, test both for routine work, and never choose by the price per token alone.”

Where each model is ahead

On an independent ranking of overall intelligence, Fable at its top setting, max, scores 65. Carbon at its top, high, scores 60. That is a clear lead, but it costs more: a task at that Fable setting averages $26.54. Fable at high scores 62, so the last few points are the costly ones.

On coding done by the AI on its own, the makers' figures on FrontierSWE v2 favor Fable: 69.9% against 61.2%. They come from the companies, not an outside test. On a new document-heavy test, Fable scores 57% and Carbon 54% at their top settings. Google DeepMind helped build it and says the top three cannot be told apart reliably, so I call it a tie. On business workflows, Carbon scores a little higher. But can you buy both where you work?

Where you can buy them

Amazon Web Services, Google Cloud and Microsoft Azure all offer Fable, Anthropic says. Vertex AI is Google's cloud marketplace for AI models. Google details paid access to Carbon through its own developer tools. It leaves three things unspecified: Vertex AI, regions (where your data is stored) and support terms (who helps when something breaks). The tester counts six companies selling Fable access. A third party reports one for Carbon. Fable also requires 30 days of data retention: the maker stores what you send for that long. I found no public word on whether the 30 days can be shortened or switched off, or on what Carbon keeps and for how long. Before you sign, I would ask each company for its terms in writing.

For hard coding and long reasoning work, I would pick Fable at high. It ranks higher than Carbon at that setting and is sold on three big clouds. At Carbon's December price, it costs only a little more per task. Until then, Carbon is far cheaper. For routine jobs, such as sorting support emails, I would test both at medium. Carbon averages $6.59 per task at the December price and scores 58. Fable costs less there and is faster. Whatever you pick, set the thinking level yourself. If the code sends none, Fable runs at high and Carbon at medium. I found no public figure for the cost of switching. Neither model runs on your own hardware, and the thinking settings differ, so I expect real work. I would pilot both on the same real jobs next to your current model, and count reruns.

Nov 13, 2026  ·  Sketchbook, p. 08

  1. Nov 17Helus says its psychedelic pill beat placebo in a depression trial
  2. Nov 17Anthropic adds text watermarking to four Claude 4 models
  3. Nov 17HHS finalizes HIPAA changes: 15-day record access and health app transfers
  4. Nov 17FBI warns attackers are using the Artex AI hacking tool against companies
  5. Nov 16Fusion launches an early funding program for defense startups

This is our sketchbook.

Two stoneware cups turned towards each other, as if in conversation

Got a good project?
Let’s work on it together.

Let’s talk →

info@ndot.io · New Jersey · Budapest · Biel