llm-anthropic 0.31 adds Fable 5.2. Max effort spent 128,000 tokens and produced no image

Simon Willison added Claude Fable 5.2 to llm-anthropic 0.31 and exposed a practical trap: max effort consumed 128,000 thinking tokens in one SVG test without producing the image. Effort is a reliability parameter, not a simple quality slider.

The plugin exposed the model in a terminal on release day

After updating, users can call claude-fable-5.2 at low, medium, high, xhigh and max effort. Willison tried an SVG pelican on a bicycle. Xhigh produced his best result, while max exhausted 128,000 thinking tokens. Fable 5 and 5.1 had completed the same max test.

A thin adapter turns an announcement into a reproducible test

The plugin does not improve the model. It shortens the path from release note to private measurement and keeps models in one environment. Teams can compare the version on their own data before rewriting production workflows.

The most expensive setting failed to deliver an artifact in one run

This was one prompt from one developer, not a systematic benchmark. It still shows an operational risk: a completed call may lack the requested file. Pricing of $8 per million input tokens and $40 per million output tokens does not prevent waste when reasoning consumes the budget.

Private evals must score output, cost and empty returns

Teams should retest each effort level, impose token limits and verify the expected artifact. Reliable completion of the same task set across repeated runs will matter more than another attractive image.

Lilith's verdict

A tap labelled max can release the most tokens and leave the bucket empty. Pipelines must inspect the finished artifact, not merely the green status beside an API call.