A concept by Patrick BusheDeveloped & drawn into this page by Claude (AI)Concept · not a product · example math
The Compute Meter
AI subscriptions cap you by the clock: a session limit that resets every five hours and a weekly limit on top. This page argues for one visible budget instead, measured in the work a request actually does, with a price shown before big jobs, rollover for quiet weeks and higher prices at busy hours.
The chain of the argument: paid AI plans sell access, then cap it with time windows → the windows already weigh heavy requests more, but you can’t see the weight before you send → unused allowance never builds up, and the wall tends to arrive in the middle of the work that matters → the real cost of a request grows with the length of the conversation it rides on → so meter the work itself, show the price of big jobs up front, let quiet weeks roll over, and price busy hours instead of walling them off.
Please read this first
This is a concept, not a product. How Claude’s plan limits work comes from Anthropic’s help center, and the caching rates from its API pricing, both linked at the bottom. The session math is an example with made-up turn sizes, and the week in C‑1 is an illustration, not anyone’s real usage. How plans weigh each message internally isn’t published.
Start here
The whole idea, in plain English
No technical knowledge needed. Read this part and you’ll understand the entire page.
AI plans don’t sell you an amount of work. They sell you access, then cap it with clocks. On Claude’s Pro and Max plans there’s a session limit that starts with your first message and resets five hours later, and a weekly limit that resets at a fixed time. The chat app, Claude Code and the desktop app all draw from the same limits.
Clocks feel unfair in two ways. A quiet morning saves nothing for a busy afternoon, because each window resets to the same full amount. And people hit the wall in the middle of the work that matters most, because that’s when they use the most.
Not all messages cost the same. A model rereads the whole conversation every time you send something. The fiftieth message in a long working session costs far more than the first, even with the discounts providers give for repeated text.
Here’s the redesign. Replace the two clocks with one meter you can see, like an electric bill or a phone data plan. Big jobs show their price before they start. Unused budget rolls over. Busy hours cost more instead of locking you out.
Quick answer
How do Claude usage limits work?
Two limits at once on most paid plans: a five-hour session and a weekly cap. From Anthropic’s help center, checked November 2026.
How it works
Plans
Session limit
Starts with your first message and resets in full five hours later.
Pro and Max. Verified Ultra accounts have no five-hour limit.
Weekly limit
Resets at a fixed time each week assigned to your account, across all models. Some plans add a separate weekly limit for Fable.
Pro, Max and Ultra.
What uses it up
The length and complexity of the conversation, the features used, the model and the effort level. The claude.ai app, Claude Code and Claude Desktop share the same limits.
All paid plans.
At the limit
Wait for the reset, upgrade, or turn on usage credits to keep going at standard API rates.
The premise · the limits already weigh your work, you just can’t see it
The help center is clear that a long, complex conversation on a big model at high effort uses your limit faster than a short question. So the plans already meter work behind the scenes. What you see is a percent bar that fills, then a wall with a reset time.
Anthropic has been moving in this direction. Usage credits let a paid plan keep working past its limits at API rates, and the Ultra plan drops the five-hour window for verified accounts. The Compute Meter takes the next step: make the unit itself visible, and let it behave like a budget instead of a timer.
5 hrs
until a Pro or Max session limit resets in full
1 week
until the weekly limit resets, at a fixed time
~185×
a 50-turn working session vs one quick question (example math)
0.1×
the input price for cached text on most Claude models
C‑1 One week, two people, same plan
Illustration · not real usage data
C‑1. Same plan, very different weeks. The steady user leaves most of every session on the table. The bursty user saved nothing by being quiet Monday to Wednesday, and loses part of Thursday to waiting. With a meter, both would draw on one budget, and the bursty user’s quiet days would count.
C‑2 What one more message really costs
Example math · 2,000 tokens added per turn
C‑2. A model rereads the whole conversation on every turn. Caching makes the reread far cheaper, but not free: turn 50 of this example still costs about five times turn 1. Long sessions are expensive because of their history, not their last message.
Add the turns up
Summed over a whole session, the gap between kinds of use gets large. Same example rates as C‑2; the cost column is in fresh-token terms, so 2,000 means “as much as 2,000 tokens at full price”.
Kind of use (example)
Tokens read in total
Cost with caching
One quick question, 1 turn
2,000
2,000
A writing session, 20 turns
420,000
~88,000, about 44 quick questions
An agent working session, 50 turns
2.55 million
~370,000, about 185 quick questions
A long agent run with big files, 100 turns of 8,000 tokens
~40 million
~5 million, about 2,500 quick questions
That’s the honest center of this piece: a working session costs as much as hundreds of quick questions, and a long agent run as much as thousands. The reply itself is left out, and output usually costs about five times as much per token as input on current Claude models, so heavy sessions are heavier than shown. People working this way aren’t light users, and a fair system should say so in numbers they can see.
C‑3 The meter on screen
The redesign · every number is an example
C‑3. One bar instead of two clocks. The grey band at the end is last month’s unused budget, carried over. Big jobs show their share of the month before they run, and offer a cheaper time; quick questions don’t interrupt you at all.
The design · five rules
A unit you can see. Each plan includes a number of units a month. A unit weighs what really costs money: fresh input, cached input, output and the model.
A price before you press go. Before a big job, such as a long agent run or a folder of files, the app shows its estimated share of the month. Small questions show nothing; they’re too small to matter.
Rollover, with a cap. Unused units carry over, up to one extra month. A light user builds a cushion and in practice never meets a wall.
Peak pricing instead of walls. Capacity is real and busy hours are crowded, so units cost more at peak and less overnight and on weekends, the way power companies price electricity. Heavy jobs drift to quiet hours on their own.
A soft landing at zero. When the budget runs out, top up, wait for the month, or drop to a smaller model, never a surprise cutoff halfway through a task. Usage credits already do part of this.
Why the clocks exist · the other side
Capacity is fixed in the short run. Data centers can’t grow by Monday morning. Short windows spread demand out without anyone having to think about it, and a meter with rollover lets a lot of saved-up work arrive at once.
Clocks are simple. “Come back in two hours” is easy to understand. A unit that blends input, cache and output has to be explained, and an estimate shown before a job will sometimes be wrong.
A meter can make people anxious. Some people would rather never watch a number tick, even if a meter would cost them less. A flat plan with generous limits may simply feel better.
The honest catch
The math is an example. C‑2 and the table assume 2,000 or 8,000 tokens a turn and the API’s caching rates. How subscription plans weigh each message isn’t published, and newer models read cached text even cheaper (a twentieth of the input price on Claude Opus 5.5), which shrinks the gaps but doesn’t remove them.
Heavy users may pay more. A meter that tells the truth about a 100-turn agent run also makes that run look expensive. The people hitting walls today may find the meter runs out just as fast.
Peak pricing only helps if work can wait. A deadline at 2 p.m. doesn’t move to 9 p.m. because it’s cheaper.
Part of this already exists. Usage credits let paid plans keep working past their limits, and Ultra has no five-hour window for verified accounts. The new parts here are the visible unit, the price before a job, rollover and peak pricing.
The whole idea, in one line
A five-hour clock can’t tell a quick question from a hundred-turn agent run. Meter the work, show the price, and let quiet weeks count.