AI Models : Prices Are Falling, But Some Tasks Cost More

AI Models : Prices Are Falling, But Some Tasks Cost More

Anthropic’s new Claude Fable 5.2 model is priced lower than its predecessor, but the cost per task can increase. According to Artificial Analysis, which evaluated the model at its maximum setting, each task costs an average of $26.54, compared to $12.99 for Fable 5.1 at the same setting.

Released on November 9, Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, compared to $10 and $50 for Fable 5.1, which Anthropic’s pricing grid now lists among its legacy models. A token is a fragment of a word used as a billing unit: a model’s invoice is therefore calculated as a price multiplied by a volume. At the maximum setting, where the model is allowed the most reasoning before responding, Fable 5.2 produced 275 million output tokens across all tests, compared to 163 million for Fable 5.1, an increase of 69% for a unit price reduced by one-fifth. The new model also scores 65 versus 49 on the same benchmark.

Price Drops, Volume Rises Faster

Consulting firm Accenture, which surveyed 750 business leaders across 17 countries, sees AI agents as a consumption multiplier, with their calls chaining on a growing context. Google observes the same phenomenon with its own model. Gemini 4 Carbon, made available to developers on November 10, is sold at the introductory price of Gemini 4 Argon, i.e., $2 per million input tokens and $10 per million output tokens. According to the Google DeepMind announcement, Carbon consumes more tokens than Argon for reasoning and extending its verification loops, and Artificial Analysis measures nearly twice as many output tokens for it.

The introductory pricing expires on November 30. Starting December 1, Carbon’s price increases to $4 and $20, and the same task will cost double for equal consumption. Argon, whose launch pricing ends the same day, also doubles. Nine out of ten leaders surveyed by Accenture say a cheaper token will lead them to expand use cases rather than reduce their spending.

Sold by the Seat, Paid by the Token

“Tokens have a significant impact on gross margin,” summarized Jean-Philippe Couturier, president of software publisher Whoz and board member of Numeum, in an interview with LeMagIT published on November 12. SaaS represents 67% of the publishing revenue of the companies studied in the Top 250 by EY and Numeum, presented on November 4, and it has historically been sold by the seat or by the license. A client who makes the agent work more still pays the same subscription, while the publisher relying on a third-party model pays its provider for each token.

In this panel, 29% of publishers have not defined a pricing strategy for AI. The premium module, the first option cited at 26%, remains a flat fee, usage-based billing attracts 22%, and value or outcome-based pricing attracts 5%. Since June 1, GitHub (a Microsoft subsidiary) customers see Copilot deducted from AI credits included in their plan, with overages charged separately. EY notes that these consumption-based costs make IT expenses harder to forecast.

The Effort Setting, the Model’s True Price

Jean-Philippe Couturier sees a solution in routing, which assigns each request to a sufficient model. A comparable trade-off also occurs within the same model. Set to the medium effort setting, which Anthropic applies by default, Fable 5.2 produces 41 million tokens for a score of 56, compared to 275 million tokens and 65 points at maximum. At this setting, it outperforms Fable 5.1 pushed to its limit, with four times fewer tokens.

In the TCS survey published on November 4 among 1,333 leaders across 25 countries, 9% say they continuously improve the value produced per token. At the other extreme, 6% admit to having limited visibility into the link between tokens consumed and results, and 31% track their costs without systematically verifying that the value obtained justifies them. In total, 37% of respondents do not regularly correlate what they spend with what they get.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x