Fable 5.2 Is Out. Start on the Cheaper Model Anyway

A new model at the top of the Claude lineup is generally available, and it is cheaper than the one it replaces. The unusual part is the advice that came with it. Anthropic's own documentation tells you to start most work on a model that costs half as much.

What shipped

Claude Fable 5.2 succeeds Fable 5.1 as the most capable generally available Claude model. Anthropic's models overview describes it as for demanding reasoning and long-horizon agentic work. A model-selection guide published on the claude.dev blog on 9 November, by Addy Osmani, positions it for the hardest technical projects in coding, finance and the sciences, and says it generates output up to 30% faster than Fable 5.1.

The price moved in the buyer's favour. Anthropic's pricing page lists Fable 5.2 at $8 per million input tokens and $40 per million output tokens, against $10 and $50 for Fable 5.1. All prices here are in US dollars and are the first-party API rates.

Haiku 5.5 to Fable 5.2 is a factor of 80 per token. Sending a tagging job to the top model wastes money, and sending a hard migration to the bottom one wastes the result.

The vendor says start lower

Anthropic's models overview is direct about it: if you are unsure, start with Opus 5.5 for most workloads, and use Fable 5.2 for demanding reasoning and long-horizon agentic work, or when your evaluations on Opus 5.5 at higher effort still fall short. The guide says the same thing in its own words, that you should start most work on Opus 5.5 at its default effort level, medium.

A vendor steering customers away from its most expensive product is unusual enough to take seriously. The reason given is that a higher price per token does not reliably mean a better result or a higher cost per task, and that there is a cheaper dial to turn first.

Effort is the dial to turn first

Each of these models can be run at an effort level, from low to max, which controls how much reasoning it spends on a request. The guide reports what that did for Opus 5.5 on one coding benchmark, SWE-bench Pro.

Two and a half times the cost for 1.4 points is a poor trade on routine work, and that is the argument for not simply turning everything up. The guide's rule is that past high, one more effort level can cost as much as a bigger model, so at that point you test the next model up at a lower effort instead.

Price per token is not cost per task

The same point runs in the other direction, and it is where Fable 5.2 earns its place. On a benchmark the guide calls FrontierCode 2.0, Fable 5.2 at high effort scored about level with Opus 5.5 at max, 45.3% against 45.5%, for roughly half the cost per task. On the hardest work, a model that costs twice as much per token finished the job for less.

The guide gives the opposite case too. On Anthropic's runs of Terminal-Bench 3.0, the previous Fable at max effort scored 58.0%, while Opus 5.5 at high scored 58.9% using half the tokens. The more expensive model did not win.

One technique in the guide transfers directly if your work has an automatic pass or fail check. Run everything at low effort and re-run only the failures at high. On a subset of the same coding benchmark, that passed about 97% of tasks for about $0.17 each, against 95.3% for $0.29 with every task at high. Better and cheaper, and it depends entirely on being able to tell which answers to send back, which is the argument in knowing which answers to trust.

It also warns about where the money goes. On one 20-problem benchmark, two problems carried 43% of the spend. Pricing a tool on your typical task and ignoring your worst one will understate the bill.

How much weight to put on these numbers

The prices are confirmed on Anthropic's pricing page. The benchmark figures are the vendor's. Some are described as Anthropic's runs or as internal benchmarks, and for the SWE-bench Pro and FrontierCode figures the guide does not say who ran them. None has been reproduced here.

They are also coding benchmarks. A quoting workflow, a support inbox or a month-end close will not behave the same way, and the general caution in how benchmarks mislead applies. Use the figures for the shape of the trade-off and get the size from your own work.

What to do with it

If you use Claude through a subscription or through software somebody else built, there is nothing to change today. The useful question is for your vendor: which model and effort level does the product run, can you choose, and will you be told when it changes. That belongs in writing, as covered in knowing which model you are paying for.

If you are building, the guide's four steps are the right ones. Pull a sample of real tasks and write an outcome check for each. Run each candidate at two or three effort levels and record the cost of each completed task. Plot score against spend. Run the winner in shadow on a slice of real traffic before switching. That is a proof of concept with a baseline, and it is a small piece of work that decides a recurring bill.

And sort the work before choosing anything. High-volume tagging and extraction belong on the cheapest model. Everyday drafting and analysis sit in the middle. The top model is for the small share of work that is hard, long-running, and expensive to get wrong. For how the previous generation was priced and limited, see our earlier guide to Claude Fable 5.

Frequently Asked Questions