The Claude Pro usage limit is eight buckets, not one number

Why the published numbers are estimates

The Anthropic Help Center says Pro gives roughly 45 messages per 5 hours and 40 to 80 hours of Sonnet 4 per week. Both numbers are ranges with caveats attached, and the caveats matter: long prompts, attachments, Opus selection, tool calls, and peak-hour multipliers all reduce the effective count. The server is not counting messages. It is counting weighted units that vary per request.

That is why two Pro users on the same plan, the same week, can have completely different walls. One person sending 30 short Sonnet questions consumes a fraction of what one person uploading a 200KB PDF and asking Opus to re-architect it does. The rate limiter sees two different utilization curves on five separate buckets, not two message counts.

The eight buckets, named

ClaudeMeter is a Rust app and we deserialize the JSON the claude.ai/settings/usage page itself fetches. The struct that holds the response is the cleanest catalogue of what the server is actually tracking on a Pro account.

Each Window is two fields: a utilization fraction and an optional ISO timestamp for when the earliest unexpired event ages out.

The two names that surprise people are seven_day_omelette and seven_day_cowork. Neither appears in the Anthropic Help Center, the public API reference, or any pricing page. They look like internal codenames for newer feature buckets that landed in Pro after the original five_hour + seven_day pair shipped. We deserialize them as plain Window structs and surface their utilization in the menu bar so when one of them is what just throttled you, you see the cause.

Read it yourself in one curl

You do not need ClaudeMeter to see this. You need a logged-in claude.ai session, your org UUID (visible in any /settings URL), and one curl with the cookie header. The response is plain JSON.

That is the whole protocol. The bars on claude.ai/settings/usage are rendered from these same eight keys. ClaudeMeter automates the loop (browser extension forwards the session, Rust binary polls every minute, popup renders) so you stop typing it by hand.

What each bucket actually represents

The bucket names in the Rust struct map onto specific kinds of activity. If you know which surface burns which bucket, you can choose where to spend your weekly budget.

Pro vs Max: same buckets, different ceilings

The shape of the response is identical across plans. What changes is the ceiling on each bucket. Pro accounts see the same eight keys Max accounts see; their utilization just hits 1.0 sooner.

How a single Pro request hits the wall

The flow is the same for every prompt. The server increments every applicable bucket, weights the increment by the prompt and the model, and the rate limiter blocks the next request the moment any bucket crosses 1.0. The 429 response does not name the bucket.

Why “limit” is the wrong word

A limit suggests one number. A budget you draw down. The server is not running that data structure. It is running eight independent rolling windows, each with its own clock, weighted by request shape, with at least two windows that Anthropic does not name publicly. “Allowance surface” or “quota envelope” is closer to what is actually happening, but those phrases would not survive a help-center pass, so the published copy collapses everything to one estimate.

The practical consequence: if you have ever been blocked when you thought you had “most of the week left,” you were probably looking at one bucket while a different bucket was the one that hit 100. The settings page surfaces this if you scroll, but the bar on top is composite-ish and it is easy to miss the secondary buckets near the wall.

What ClaudeMeter does with this

One brew install, one browser extension that forwards the session cookie, and a Rust binary that polls the same eight-key endpoint every 60 seconds. The menu bar popup renders every bucket the server returned, with its utilization rounded to a percent and its resets_at converted to a relative timestamp. When Anthropic adds, renames, or removes a field, the strict deserializer fails loudly and we ship a release. The numbers in the popup match the numbers on /settings/usage exactly because they are the same numbers.

ClaudeMeter does not estimate, does not count tokens, does not read your local Claude Code logs. It surfaces the server's utilization fractions on the eight buckets the rate limiter actually checks. Free, MIT licensed, no telemetry, single HTTPS request per minute to claude.ai using your existing session.