Anthropic released Fable 5.2 and Mythos 5.2 today, expanding the available reasoning tiers for their latest model series. The developer Simon Willison updated his llm-anthropic tool to support these versions, allowing users to generate images with varying levels of cognitive effort. He tested the system by asking models to create SVGs of a pelican riding a bicycle across five distinct reasoning settings. The results showed that the xhigh tier produced the most coherent image, displaying correct leg placement and a functional bicycle frame, though minor alignment issues remained in the pouch area.
The max tier failed completely, consuming 128,000 tokens during internal processing before stopping without output. This mirrors a specific failure mode seen previously with Opus 5.5, where excessive reasoning caused token exhaustion. The issue appears specific to the Fable 5.2 release, as earlier versions like Fable 5 and Fable 5.1 successfully handled maximum effort requests for the same prompt. This highlights a regression in efficiency or token budget management for the new maximum reasoning level.
- The
maxtier consumed 128,000 tokens before failing. - Earlier Fable versions handled maximum effort requests successfully.
- The
xhightier produced the best visual result.
