Anthropic Releases Claude Fable 5.2: Twice the Price of Opus 5.5, but Cheaper Than 5.1

Overnight, Anthropic released Claude Fable 5.2 and Mythos. Everyone has already recapped the specs and benchmarks, so it is more useful to look at the details that are easy to miss: what exactly was built into the training, what it costs in practice, and whether it is worth moving off Opus 5.5 at all. Below is a breakdown based on observations by the author of the Tips AI channel. A caveat: this is his interpretation, not the company's official position.

A Model with an AI R&D Ceiling

The main observation: Fable and Mythos are the first models that Anthropic deliberately capped at the Opus 5.5 level in AI R&D tasks during training, that is, in research and development of AI itself. Put simply, in other respects the model may be stronger, but it should not help build the next models any better than Opus 5.5 can.

There are two versions of the reason. The first, a pragmatic one: this is the strongest protection against distillation. Distillation is when a competitor collects the answers of a strong model and trains its own, cheaper one on them.

If a model is intentionally weaker specifically in ML tasks, such a "student" gets less benefit. The second version is more serious: Dario Amodei may fear recursive self-improvement. A model speeds up the creation of the next one, that one speeds up the creation of a new one, and at some point people lose control of the process.

The post offers no confirmation, and both versions remain guesses. But the very fact that a lab deliberately cuts capabilities in one area is a notable signal for the industry.

Prices and Limits: Official and Real-World

On the price list, the picture is as follows (input and output price respectively):

• Opus 5.5 — $4 / $20
• Fable 5.2 — $8 / $40
• Fable 5.1 — $10 / $50

Fable 5.2 is cheaper than the previous version but twice as expensive as Opus 5.5. In addition, a separate limit applies to Fable — half of the main one.

On the $200 subscription the difference is more tangible. By the author's estimate, with Fable you get roughly $3,000 of usage, and with Opus 5.5 — $15,000. In other words, per subscription Fable gives five times less volume. The author admits, though, that he hasn't hit limits in a long time. Moreover, Anthropic gave subscribers from one to three limit resets specifically for Fable: 1 reset for subscriptions under 6 months, 2 for 6 to 12 months, 3 for more than a year. The author, as a longtime user, received three.

What to Do in Practice

First: don't switch everything over. The author notes that Opus 5.5 still handles all of his tasks, and he can't imagine what Fable would need to be able to do for the difference to become noticeable. The approach is sound: if your pipeline runs reliably on Opus, changing models will double your bill with no guarantee of a gain.

Second: run Fable 5.2 on your own cases. Take 20–30 real tasks where Opus currently makes mistakes or needs several attempts, and compare the cost of a successful result, not the price per token. If Fable solves on the first try what takes Opus three iterations, the double price may pay off.

Third: read the official Fable 5.2 prompting guide in Anthropic's documentation. Recommendations change with each new model, and old prompts may perform worse than they could.

Fourth: if your product depends on ML tasks (generating training data, automating experiments, code for training models), keep the stated limitation in mind. This is exactly where you shouldn't expect gains over Opus 5.5.

And finally, on limitations. The source is an author's post on Telegram, and the calculation of $3,000 and $15,000 on the subscription is a personal estimate, not official figures. The post does not give Mythos's price, and the details of the AI R&D limitation should be verified on the release page on Anthropic's website. For now this is more a reason for careful testing than for mass migration.