Anthropic releases Fable 5.2: near-Opus performance at almost half the cost per task

An AI agent can spend hours fixing a bug, running tests and rewriting its own patch. For the team, the question quickly becomes financial: how much does it cost to produce a change that a developer will actually accept? Anthropic has released Fable 5.2 and says that, on one coding benchmark, the model came close to Opus 5.5 at almost half the cost per task. That is a reason to test your agent's settings, especially if spending on extended reasoning is already making a noticeable dent in the budget.

Where the savings come from

On FrontierCode 2.0, Fable 5.2 with the High setting scored 45.3%, while Opus 5.5 with Max scored 45.5%. The benchmark assesses whether maintainers of open-source repositories would accept the changes proposed by the model. According to the company, Fable's cost per task in this comparison was roughly half as high. The new model's API costs $8 per million input tokens and $40 per million output tokens. The figures come from Anthropic's announcement.

It is useful here to distinguish the price of a token from the cost of a result. The former appears in the pricing table. The latter depends on how many times the agent tries to fix the error, how much context it reads and how much work it leaves for a human. A cheaper run can easily end up costing more if an engineer then spends half a day restoring broken behavior.

The reasoning setting also becomes part of the choice. Comparing models only at their maximum settings makes for a convenient leaderboard. For a product, it is more useful to find a setting that delivers acceptable quality within an affordable budget. Anthropic's result shows why that experiment is worthwhile, although no one can promise a particular team the same savings in advance.

How to test it on a familiar bug

Imagine a subscription service: after a repeated request, a customer sometimes receives two identical charges. The agent needs to investigate how retries are handled, fix the code and add a check. An elegantly written explanation achieves nothing here if duplicate charges keep appearing.

To compare models, prepare the same repository snapshot, bug description and available tools. Set identical time and spending limits. Then check whether the patch resolves the original problem, whether existing tests pass and whether behavior changes in related scenarios. It helps if the reviewer initially assesses the changes without seeing the model's name: that reduces the temptation to forgive a favorite's mistake.

One bug is not enough. Assemble a small set of previously solved tasks: fixing a regression, extending an API, changing a data schema and finding the cause of a slow query. For each one, record the cost of runs, the number of retries and the time spent on human review. The primary metric is cost per accepted fix. Track the task success rate and completion time alongside it; otherwise, savings could conceal a decline in quality or speed.

For a product manager, this is a way to test a specific hypothesis: can you handle more complex requests on the same budget? For a founder, it is an opportunity to assess a feature's cost before promising users an unlimited assistant.

What the announcement does not establish

The published comparison has a significant limitation: the results were presented by the vendor itself. Your repositories, tools, instructions and acceptance criteria may differ substantially from those used in the benchmark. The difference between 45.3% and 45.5% also does not, by itself, establish that the models are equivalent across all engineering tasks.

Switching models across the entire product immediately would therefore be premature. A sensible first step is to run the saved tasks, then give the new configuration a small share of real work with the usual review process. If the gains hold up, usage can expand gradually while retaining the option to revert to the previous setup.

The release's practical value lies in a testable promise to reduce the cost of complex work. Each team needs its own answer to a simple question: how much money and time does it take to produce a fix that passes review and makes it into the main codebase?