Anthropic makes Claude Fable 5.2 up to 30% faster and cuts token prices

It is a familiar situation: an agent enthusiastically changes dozens of files, yet the migration still fails its tests. Anthropic has released Claude Fable 5.2 and Mythos 5.2 for tasks like these. The practical question is whether the new model can carry complex work through to an accepted patch at a reasonable cost. Test that in your own repository: early independent measurements show strong results, but also warn of long task completion times.

What changes for developers

According to Anthropic's announcement, Fable and Mythos use the same base model. Fable has stricter restrictions on cybersecurity and biology tasks; organizations gain access to Mythos through vetting programs. Describing Mythos as a model with no safeguards whatsoever is therefore too sweeping: it has a different access framework for certain applications.

The company claims generation is up to 30% faster than in version 5.1. Standard pricing is $8 per million input tokens and $40 per million output tokens. An optional fast mode promises generation up to 2.5 times faster, but doubles the rates to $16 and $80, respectively. These figures concern generation, not a promise to finish every engineering task proportionally faster.

Suppose a team is migrating a service to a new version of a library. The agent needs to find incompatible calls, update wrappers, run tests and fix the resulting issues. Faster code generation helps with only part of that process. If the model spends a long time reasoning, repeatedly calls tools or makes unnecessary changes, the gain in output speed can disappear. That is why comparisons should cover the entire workflow.

First place does not eliminate the wait

In Artificial Analysis, Fable 5.2 configured with the highest reasoning level scored 65 points and took first place in the Intelligence Index. However, the average completion time for an index task was 47.5 minutes, at a cost of $26.54. These figures describe a specific test suite and configuration; they cannot be applied directly to your service.

For a product manager, that is a significant caveat. The model could be useful in a background agent assigned a difficult task whose result is checked later. In an interface where users expect an immediate response, the same configuration needs separate latency testing. A high ranking alone does not settle the question of user experience.

Vals AI also ranks the model first in its overall index, with 71.31% and a stated margin of error of ±0.96 percentage points. It scores 93.77% on Vibe Code Bench, and 83.27% and second place on the code migration test. The picture changes considerably across specialties: the model ranks 20th in the legal index and 27th in the tax agent benchmark. These results do not establish a universal winner for every workflow.

How to test its value on your own project

Choose several completed tasks with known correct behavior and acceptance criteria: a dependency migration, a cross-module bug fix, or an API change that includes updating clients. Give the new model and your current model identical starting conditions. Measure time until tests pass, spending across all attempts and engineering time spent on review.

Test fast mode separately. Double the price could be worthwhile if the wait really shrinks and quality holds up. If builds and tests account for most of the time, the effect may be modest. This is a hypothesis to measure, not a conclusion you can draw from the announcement.

For a founder, the useful metric here is the cost of completed work. A more expensive model is justified when it reduces rework or unblocks a task that was costing the team days. Fable 5.2 deserves that trial. A decision on broad adoption is best made after checking accepted changes, latency and the full execution bill.