Fable 5 – Median thinking declined in August

Lon Lundgren reports that Fable 5’s median thinking usage, the amount of reasoning text the model produces before answering, fell from July to August across five ways of counting or weighting the data. The chart covers 6,921 turns and 36,374 invocations at settings labeled “xhigh” and “max effort.” The decline ranges from 21.9% with turns weighted equally to 50.6% when measured per invocation.

The useful finding is that the direction holds across these summaries, while its size depends on how usage is counted. They are different views of one observational dataset, not independent replications. The chart gives only aggregate counts, so it does not show whose usage was sampled, whether tasks were comparable across months, or whether answers got worse. It cannot explain the decline or say that every user was affected.