Anthropic hasn't changed Claude's limits since 30 Jul 2026.

A small fixed prompt is run repeatedly on an otherwise idle Max 20x account until Anthropic's own usage meter ticks from one whole percent to the next, twice. The tokens spent between the two ticks are one percent of the 5-hour window. That runs twice a day, rotating across Sonnet 5, Opus 5 and Fable 5.1.

The split between input, output and cache tokens comes from real working sessions on a second Max 20x account, joined against the same usage meter. Effort figures come from one calibration task run at every effort level on every model.

This is one account running one workload mix; your split of input, output and cache tokens will differ. The usage meter reports whole percent, so every probe carries about one small prompt's worth of rounding error. Effort figures describe one calibration task shape, not your actual work. Pro and Max 5x are scaled from Max 20x using Anthropic's published plan ratios, not measured directly — the passive-observed ratio between Max 5x and Max 20x on Jonathan's own account history came out at roughly 2.5x, well below the published 4x, most likely because some of that usage happened off this machine and wasn't captured by the probe.