Claude Fable 5.2 is 20 percent cheaper per token than its predecessor, but in an independent test of professional tasks it was 31 percent more expensive per task. If you budget AI work on the price per token, your numbers look better than they are.
What happened
Anthropic released Claude Fable 5.2, its new top model. Two serious newsletters looked at it from different angles.
Latent Space looked at the cost. The price dropped from 10 to 8 dollars per million input tokens and from 50 to 40 dollars per million output tokens. Even so, SemiAnalysis paid 88.15 dollars per task on GDPx, a test of professional tasks, against 67.16 dollars for Fable 5.1. The logs explain why: where earlier models tried about twenty test cases, Fable 5.2 sometimes wrote hundreds and spun up extra subagents to check its own work. At the highest thinking level (max), Artificial Analysis arrives at about 26.5 dollars per task, roughly 2.6 times as much as Opus 5.5. Several evaluations also found that the highest levels score worse than the high level.
Ethan Mollick looked at the quality. He gave the model hard assignments across very different fields, from a business plan to an academic paper and a website with hundreds of 3D worlds. On many tasks he found the work to be at the level of a human expert. But the model now fails at something different than before: not at made-up facts, but at taste and judgement. A course website was accurate but dull, a paper grew to 182 pages because nothing was ever cut, and a rule book the agents wrote themselves grew to 204 sections and 87,500 words in twelve days. Most important: the model did not notice these problems itself. Mollick calls this "jagged": very strong, but not everywhere.
What this changes for you
The two sources complement each other. Latent Space shows that more thorough work costs more tokens, Mollick shows that all that extra checking does not automatically produce better work. He also notes that his token use was not metered during the test period, so his experiments say nothing about what they would have cost.
For a managing director or IT manager, that means:
Budget per task, not per token. Take three tasks you actually hand to AI, such as drafting a quote or summarising a report, and measure what each one costs per run. A lower token price can still mean a higher cost per task.
Do not switch on the heaviest model and the highest level by default. Anthropic itself advises starting at medium or high and keeping the highest levels for the hardest questions. For well-defined, repeatable work, a lighter model is often enough.
Keep a person on the final judgement. The model checks facts thoroughly, but does not ask whether something is too long, too repetitive or worth the effort. That stays the job of someone who knows what the result has to deliver.
Check the data agreements. According to Latent Space, Fable 5.2 retains data by default so that Anthropic's safety checks can run; zero data retention is only available to eligible enterprise customers. Do not put customer data into a chat before you know what happens to your data.
How it works under the hood
Fable 5.2 has five thinking levels: low, medium, high, xhigh and max. At a higher level the model checks more, with more test cases and subagents, and that costs tokens. According to Anthropic, Fable at high matches Opus 5.5 at max on a coding test for about half the cost per task. Cursor advises starting with the default level and not carrying over settings from Fable 5.1, because the levels behave differently. The context window is 1 million tokens. If you use the model in your own application, measure cost and quality per task with your own examples instead of relying on public rankings.
We work with Claude every day and budget AI work per task. We only roll out a new model widely once we have seen its cost and quality on our own work.
This news touches chapters 6, 19, 23 and 24 of the course AI(l) or nothing by Newway. If you follow the course, this subject returns there at a calmer pace.
What you can do with this
Want to know what this means for your business? Plan an introduction: half an hour, no sales pitch, but an honest answer to the question where to begin. Rather read on first? The course AI(l) or nothing is free and starts with the basics.
Written in response to publications by Latent Space and One Useful Thing by Ethan Mollick.
