Claude Fable 5.2 carries a lower price tag than its predecessor: 8 dollars per million tokens for entry and 40 for exit, compared to 10 and 50 for Fable 5.1, which the Anthropic grid now ranks among its old models. Artificial Analysis put the new model, released on November 9, through its battery of evaluations, tuned to maximum effort. Each task costs him an average of $26.54. Fable 5.1, at the same setting, came in at 12.99. The newcomer responds better, with an index of 65 against 49. He also writes a lot more.
The price goes down, the volume goes up faster
A model pays by token, this fragment of a word which serves as the unit of account for everything he reads and writes. Its invoice is a price multiplied by a volume. At the maximum setting, the one where the model gives itself the most thought before responding, Fable 5.2 produced 275 million output tokens across all tests, compared to 163 million for Fable 5.1, or 69% more for a unit price down by a fifth. The Accenture firm, which surveyed 750 business leaders in 17 countries, sees agents as a consumption multiplier, because their calls are linked in a context that lengthens.
Google recognizes the phenomenon for its own model. Gemini 4 Carbon, opened to developers on November 10, is sold at the introductory price of Gemini 4 Argon, $2 per million tokens in and 10 out. In its announcement, Google DeepMind specifies that Carbon consumes more tokens than Argon to reason and extend its verification loops; Artificial Analysis measures almost twice as much output. The introductory price expires on November 30, a deadline that Google did not set at the end of October. From December 1, the price increases to 4 and 20 dollars, and the same task will cost double for equal consumption. Argon, whose introductory price ends the same day, is also doubling.
Nine out of ten executives surveyed by Accenture say a cheaper token will lead them to expand use cases rather than reduce spending. Buyers won’t put on the brakes on their own.
Sold at headquarters, paid by token
“Tokens have a significant impact on the gross margin,” summarized Jean-Philippe Couturier, president of the publisher Whoz and administrator of Numeum, elected to the college of software and platform publishers, in an interview with LeMagIT published on November 12. SaaS represents 67% of the publishing revenue of the companies studied by the EY Top 250 and Numeum, presented on November 4, and it has historically been sold by headquarters or license. A client who makes the agent work more pays the same subscription. The publisher who relies on a third-party model pays each token to its supplier.
In the panel, 29% of publishers have not defined a pricing strategy for AI. The premium module, the first option cited with 26%, remains a package that moves the problem up a notch without resolving it. Billing by use attracts 22%, and pricing by value or result, the closest to a price per task, 5%.
Since June 1, customers of GitHub, a Microsoft subsidiary, have seen Copilot deducted in AI credits included in their package, and the excess is payable. EY notes that these consumption-indexed costs make IT spending more difficult to anticipate. For a CIO as for a financial director, the AI budget line ceases to be a number of licenses.
The effort adjustment, real price of the model
Jean-Philippe Couturier sees the solution in routing, which entrusts each request to the sufficient model. The project has been completed by certain publishers and is a priority for everyone, he said. An identical trade-off occurs within a model. Set to average effort, the one that Anthropic applies by default, Fable 5.2 produces 41 million tokens for an index of 56, compared to 275 million and 65 points at maximum. At this setting, it does better than Fable 5.1 pushed to the max, with four times fewer tokens.
Google itself intends Carbon for the toughest tasks and judges that Argon remains the right choice for the majority of software development and intellectual work. Accenture estimates that about 90% of the tasks it has studied do not require a cutting-edge model.
In the TCS survey published on November 4 among 1,333 executives in 25 countries, 9% say they are continuously improving the value produced by token. At the other end, 6% recognize limited visibility on the link between tokens consumed and results, and 31% track their costs without systematically checking that the value obtained justifies them. More than a third of respondents, 37%, therefore do not regularly compare what they spend with what they obtain. And consumption is still ahead of them, with only 5 to 7% describing their technical environment as fully scaled.
The list price of a model, as we noted it in October, no longer indicates what a function costs. The unit that counts is the task. The 29% of publishers who have not decided on their schedule will not be able to build it on the public price of Anthropic or Google.
Consumption that suppliers do not publish
Measuring requires rare data. Anthropic and Google promote their models in cost per task, but in relation to another model and without giving the number of tokens consumed. Artificial Analysis counts output tokens on evaluation tasks, far from payroll or shift planning. The TCS, EY and Accenture surveys are declarative and report what executives say, not their invoices. The only reliable counter is the one that a publisher installs in its own product, function by function.
A publisher who sells his AI to headquarters and pays for it in tokens runs an all-you-can-eat buffet from which he buys the meat by weight. On November 9, the kilo fell by a fifth. The plates come out of the kitchen more garnished, and the guests serve themselves again.
Our articles will be prioritized in Google’s “Featured” section.

