Twice the price per token, not necessarily per task
Let's start with the price lists, because they are the easy part. Opus 5.5 costs $4 per million input tokens and $20 output. Fable 5.2, released on November 9, costs $8 and $40. Exactly double.
The less easy part is that nobody pays per token. You pay per task: how many tokens the model reads and writes, how many times it rereads the context, how many attempts it makes before finishing, and whether it finishes at all. A model that costs twice as much per token but completes the job in half the steps costs the same. One that costs twice as much and stops halfway costs infinitely more, because a person then redoes the work.
That is why the comparison between Fable 5.2 and Opus 5.5 cannot be settled by looking at prices. You need to understand where the two models are equivalent, where they aren't, and how much effort matters. What's new in the release is in the article on Claude Fable 5.2; here I focus on the choice.
Where they are equivalent
On coding and office automation the two models are closer than the price suggests.
On FrontierCode 2.0, agentic coding, Fable 5.2 at xhigh effort reaches 49.1%, Opus 5.5 at maximum effort 45.5%. On OSWorld 2.1, which measures the ability to use a computer the way a person would, 89.3% against 87.2%. On FrontierSWE v2, 69.9% against 62.3%.
These are real differences but not dramatic ones. On a well-defined development task, on an automation that fills in forms and moves data between applications, on document analysis, Opus 5.5 does nearly the same job at half the price. For a broader view of the Claude lineup there is the Sonnet, Opus and Haiku comparison.
Where Fable pulls ahead
The picture changes when the task becomes long, open-ended and hard.
On Terminal-Bench-Science, agentic scientific research, Fable 5.2 reaches 74.6%, Opus 5.5 58.7%. On Humanity's Last Exam with tools, 74.2% against 67.7%. On reading charts without tools, 73.4% against 64.4%. On RoboDojo, a robotics simulation, 64.9% against 39.4%.
The case that makes the point better than any table is the formalization of Fermat's Last Theorem. About a thousand Fable 5.2 agents completed it in 30 hours. Opus 5.5, with the same setup, was at 26% after 18 hours and then barely moving. It is not about the quality of a single answer: it is the ability to stay on course through work that lasts days without getting lost.
In business, tasks like this exist, but there are fewer of them than people think. A technical analysis that runs across hundreds of documents and has to reach a verifiable conclusion. An agent working for hours on a large codebase. Real research and development. There, the price difference is justified. Elsewhere, rarely.
Not sure which processes really need Fable?
30 minutes to discuss your specific case.
Effort changes the comparison
There is a wrong way to compare the two models, and it is putting them side by side with default settings. The Fable 5.2 API defaults to high effort, while Opus 5.5 in the apps defaults to medium. You are comparing two different things.
The Fable 5.2 system card contains a useful table on FrontierCode 2.0. At low effort the model solves 34.2% of tasks at $2.10 each. At medium, 37.6% at $2.48. At high, 45.3% at $6.72. At xhigh, 49.1% at $14.19. At max, 45.7% at $44.70.
According to Anthropic, at high effort Fable 5.2 matches Opus 5.5 at maximum effort at roughly half the cost per task. If that holds, on some workloads Fable at moderate effort is better value than Opus pushed to the limit.
And there is a lesson that applies to both models: maximum effort is rarely the right choice. On this benchmark max costs three times xhigh and does worse. Anthropic's own documentation recommends starting at high, moving to xhigh only where you have measured a gain, and keeping max for measuring the model's ceiling.
What independent measurements say
Anthropic's numbers should be taken for what they are: tests chosen and run by the company that makes the model. The first external measurements tell a story less favorable to Fable on cost.
On Terminal-Bench 4.0 Vals AI measured $14.56 per test with Fable 5.2 against $12.04 with Opus 5.5. 21% more, which on that benchmark is a modest premium.
On CursorBench 4.0, both at maximum effort, Cursor reports $32.79 for Fable 5.2 against $13.43 for Opus 5.5. More than double. With Fable at high effort the cost drops to $9.70.
My reading: Fable's premium on cost per task ranges from small to large, depending on the task and above all on effort. There is no number that holds for everyone, which is exactly why the choice is made on your own tasks and not on someone else's tables.
A practical rule
The rule I use with clients is simple to state.
Opus 5.5 is the baseline. Every process starts there, at the lowest effort that gives acceptable results.
Fable 5.2 comes in only where Opus 5.5 doesn't finish. Not where it finishes a bit worse: where it stops, slips on long sequences or needs a person to step in. Finding out takes three or four real tasks, run on both models with the same prompt, measuring two things: how many get there without help and how much they cost.
One technical note that matters for anyone building agents. Fable 5.2 thinking blocks are read only by Fable 5.2 and Mythos 5.2. If a router moves a conversation from Fable to Opus midway, the accumulated reasoning is lost and the model restarts with less context than you think. Moving from the smaller model to the larger one works, the reverse doesn't. Decide the model at the start of the task, not in the middle. I wrote about designing this kind of architecture in autonomous agents with the Agent SDK.
For a European company
Two concrete considerations before deciding.
The first is about data. Fable 5.2 requires 30-day retention for safety monitoring and is not available under zero data retention without Anthropic's express authorization. If your contracts or policy require zero data retention, the choice is already made, and it isn't Fable.
The second is about subscriptions. At launch Anthropic included Fable in the weekly limits of Pro and standard Team plans, as a promotion. It is the cheapest way to get a feel for it on real tasks before spending on the API, but remember the promotion is not the price list.
On budget, the right question is not which model is best. It is on what share of the work the best model changes the outcome. In most of the companies we work with, that share is small. For the full cost picture there is how much Claude costs for businesses.