Anthropic last increased Claude's limits by 20% on 22 Sep 2026.

Throughout the day the tracker reads two things off each Max 20x account it watches (4 accounts with usable readings today): the Claude Code transcripts of the work actually done on it, and that account's own usage meter. Between any two meter readings it knows how far the meter moved and which tokens were spent moving it, and that gives a price for one percent of the five-hour window. Readings from every account are pooled by day. Nothing is run to produce the limit figures. They come out of ordinary working days.

The meter does not treat every token the same. Cache reads cost nothing against it. Input, output and cache writes are charged at Anthropic's list price. So every reading here is an API-dollar value per percent of meter. The token figures are not converted from it: they are the tokens those same stretches carried, counted per percent of the meter. Headless (claude -p) work moves the five-hour meter faster, so on five-hour figures its tokens count at the measured factor shown below the window. The effort figures come from one calibration task, run at each effort level on each model.

The weekly limit is measured the same way on the seven-day meter: the tokens behind each 1% it moves, times 100. Windows per week is the weekly figure divided by the window. A change is dated to the day it lands rather than averaged into a calendar week.

Token figures depend on the token mix. Because cache reads cost nothing against the meter, cache-heavy work gets far more tokens per window than cache-light work for the same dollar value. The split shown is one account's real mix; yours will differ, and the dollar figure is the one that carries across. The meter reports whole percent, so each reading carries up to a percent's worth of rounding.

The tokens and dollars per window for Pro and Max 5x are scaled from Max 20x by the credits table: 1 : 6 : 20 per five-hour window and 1 : 8.33 : 16.67 per week (she-llac.com/claude-limits, as of 25 Jan 2026). The weekly window counts are measured on Max 20x. The Max 5x history (Jun–Aug) is measured; its current figure is inferred from Max 20x since the 13 Sep cut, and Pro is inferred the same way.

The effort figures describe one task shape, run the number of times each cell of the matrix above prints. On Sonnet the spread between runs is wider than the gap between low, medium and high, so read those three rows as roughly equal rather than in order.

27 harness runs are excluded from the stretches behind these figures. Fable 5.1's credit rate: measured by one fit pooled across every account, 1.8629 to 2.1181 credits per input token.

This method is only as good as what it can see. Work done away from the machine being read moves the meter with no transcript to match it. Where that is obvious, because the meter moved with no transcripts at all, the stretch is dropped. Partial use elsewhere is not obvious, and it reads as a cheap day.

The current figure does not wait for a week to finish. It is measured from the seven-day meter since the last confirmed change, so a step is dated to the day it landed rather than blended into a week's average.

Every number on this page comes from the JSON at /data/claude-usage.json.

Source code and raw data: github.com/jonathanavis96/claude-usage-tracker.

Method: median over the pure-opus clean stretches of the tokens the stretch carried per 1% of the five-hour meter, times 100, taken per token class and over all four classes together -- `all` over the current side of the cut alone (`current_method`), `per_class` over the whole cluster, and both sides published beside each. The interval is the spread of the readings themselves -- per account its own lowest and highest per-stretch value, pooled the union of the account intervals -- and not a confidence interval. No credit rate and no class weight enters the pure-opus figure: it is the stretch's own token counts over its own meter movement. Every other family is a conversion of it at the two families' rates (`conversion`), and a family whose rate is a status sentence publishes no number. A regime after a known-date change is measured the same way in its own regime family (`measured_family`, `current_method`), and that family's row is its direct reading. Measured to 21 Sep 2026.

Opus 5.5 rate, as of 9 Oct 2026: Opus 5.5 is valued at its list price, 0.8x Opus; the fitted 0.762x Opus is not used, because the joint fit at its first use (span test: the new family's share of the work varies too little to separate its rate from the limit change: the rate's 95% interval runs 0.82x to 1.33x its base family's, wider than 1.5x end to end or against a search bound) cannot tell that rate apart from a five-hour limit change at that instant. The fitted rate takes over once that fit separates..

34,957,676 credits per 5-hour window (28,903,750 to 39,904,825), n=12, pure-opus stretches, on 2 accounts. Max account 1, Max account 4 contributed no clean pure-opus stretch.

0.0% input·0.4% output·97.0% cache read·2.5% cache write

Split: history/passive.json `split`, the watched accounts' own token-class shares.

Cache-normalised at that split, over a median session of 1.9M tokens.

A week currently holds about 4.3 five-hour windows, measured from 4 accounts.

Tokens per week chart: each line is the limit itself, held flat between changes -- a step means a measured change, and nothing else on the chart moves. Solid and shaded: selected plan. Grey: the others. Dashed: inferred from another plan by the measured plan ratio, not measured on this one. Red: a measured change.

Each watched account's own figure: Max account 1 4.47 (179 readings), Max account 2 4.64 (134 readings), Max account 3 4.96 (56 readings), Max account 4 4.52 (31 readings).

Each line is the limit itself, held flat between changes: a step means a measured change. Solid: selected plan. Grey: the others. Dashed: inferred from another plan by the measured plan ratio, not measured on this one. Pro and Max 5x are currently inferred from Max 20x. Reading: one five-hour window on one account, hollow where the seven-day meter moved under 5%, this chart's own threshold for drawing a reading hollow. Weekly: a calendar week of readings pooled, the whisker its rounding interval, hollow while the week is in progress.

Five-hour windows per week: 6.3 before, 4.8 after.

Each account's own meter either side of 14 Sep 2026, in credits per 1% of the five-hour meter.

Max 20x weekly figures are measured from 4 accounts. The Max 5x history (Jun–Aug) is measured; its current figure is inferred from Max 20x since the 13 Sep cut, and Pro is inferred the same way.

Basis: credits_table. Credits per five-hour window, Pro : Max 5x : Max 20x: 550,000 : 3,300,000 : 11,000,000. Credits per week: 5,000,000 : 41,666,700 : 83,333,300. Measured confirmation: Max 5x over Max 20x 1.66, Max 5x Jun-Aug over Max 20x 19 Aug-11 Sep. Source, as of 25 Jan 2026: she-llac.com/claude-limits.

Measured against the reference table: this tracker's measured five-hour windows per week against the reference table's, per plan, for the last regime that ended before the 14 September weekly change. Cut at 14 Sep 2026.