Claude Code usage limits by plan: how many tokens is Max 20x?
Published 29/10/2026
Claude Code plans don't give you a number of tokens. They give you two meters, each running from 0 to 100%, and how far a request moves them depends on the model and on the kind of token it spends. So the honest answer to "how many tokens?" is an exchange rate.
Model a5v3, last checked against a live meter on 29/10/2026, on a Max 20x plan. More: plan usage burn data.
The two windows#
- The weekly meter is the budget. It resets once a week.
- The five-hour meter is a speed limit inside it. A full five-hour window is about 25% of a week, so you can empty it about four times before the week runs out.
What 1% of a Max 20x week buys#
| What you spend | Opus tokens per 1% | Sonnet tokens per 1% | Those Opus tokens at API prices |
|---|---|---|---|
| Output | 660k | 1.1M | about $13 |
| Input | 3M | 5M | about $12 |
| Cache read | 170M | 290M | about $35 |
| Cache write | 4.6M | 7.7M | about $37 |
Real work mixes all four, so read these as exchange rates rather than an allowance. Output is the dearest token, and Sonnet stretches the week further than Opus: the same week buys about 110M Sonnet output tokens against 66M for Opus.
Other plans#
We measured Max 20x. If the plan names mean what they say, Max 5x is about a quarter of these figures, but we haven't checked that against a meter.
The limits move#
On 14/09/2026 the limits changed, and the same work now reads higher on the meter than it did before. Pages of hand-typed numbers go stale. These figures are regenerated whenever the model is refitted, and the date at the top says when it last matched a live meter.
Where ADE fits#
ADE reads the same two meters, shows each engine's week against the share you set for it (ade status --line, ade cockpit), and parks work rather than waking a model when an engine is over its share. See Day to day.