Claude Pro vs Max vs Ultra: Which Plan Is Worth It, Measured

The question is not which Claude plan is best. It is how many points of a weekly allowance your week burns, and nobody publishes that number, so I measured mine. Claude sells four tiers for Claude Code: Pro at $20, Max at $100, Ultra at $200, and Team seats at $20 or $100 a head; the pricing page describes them as multiples of each other and never says what the base multiple is. Since September 2026 a governor on this machine has polled the usage endpoint of one $200 account every five minutes while I steered Claude Code from a terminal and one or two unattended workers shared the account. That gives eight complete weekly windows, seventeen five-hour windows from the last week the five-hour window existed on the account, and one real cap. Scaled by the vendor’s own multipliers, those weeks say which tier would have held them, and the picker below applies the same arithmetic to yours.

TL;DR — Buy Pro if the agent gets about four busy hours a week from you or less, and you are happy to wait for a reset. By the measured rates a busy hour at the keyboard is about a quarter of a Pro week, and every one of my eight weeks would have capped on Pro, most of them by Thursday. Is Max worth it? Buy Max if you work with the agent most days and nothing runs while you sleep: five of my eight weeks would still have capped on Max, but three fit, and one always-on worker sits at about 84% of a Max week. Buy Ultra if you run anything unattended or you steer for thirty hours a week; on my account it held every week at 30 to 80%, and the five-hour window stopped existing on October 31, which is the tier’s one structural difference. Team Premium seats, billed annually, cost the same as Max per person and buy roughly the same allowance with one bill; pay per token wins below about five busy hours a week with no plan in hand, and nowhere else.

Which plan fits you?#

If the picker does not load, these are its three questions, and the sections below give the rates behind them: how many busy hours a week you spend at the keyboard with the agent (0 to 60); how many agents you keep running unattended, all week (0 to 4); and how many people would share the bill on Team seats (1 to 10). It answers with the share of each tier’s weekly allowance your week would use, the hour the cap would land where it would, and what the same week costs at list price through the API.

What was measured#

One Claude account on the $200 tier, from one machine, September 16 to November 12, 2026. Four kinds of evidence:

• The usage meter. A governor polled the vendor’s usage endpoint every five minutes and kept each window’s reading, its reset time and how many unattended workers were alive: 5,000 readings of the weekly window across nine windows, 813 of the five-hour window across its last seventeen.
• The transcripts. Every Claude Code session on the machine from July 9 to October 31, read once for the earlier Codex vs Claude Code study: 3,000 terminal sessions on Anthropic’s models, 700 of them an hour or longer, with the CLI’s own list-price estimate per session and every usage-limit error.
• The caps. Three alerts from August, when the account was driven hard by hand, and one real weekly cap on October 12 and 13, which the transcripts record as 118 usage-limit errors.
• The pricing page, read twice on 2026-11-11.

Three caveats before a number. First, the fleet was throttled: the governor held the unattended side to two Claude workers, and most of the account’s use was a person steering, so the weekly peaks are a picture of one developer plus a small fleet, not of a tier’s ceiling. Second, the multipliers are the vendor’s. The pricing page says Max is five times Pro and Ultra ten times, and never says what Pro is, so the Ultra row below is observed and the Pro and Max rows are modelled from it. Third, the account changed shape during the study, which turned out to be the most useful thing it did.

What each tier buys#

The $200 tier is called Ultra on the pricing page as it reads today. Max 20x is the old name of that tier: when I subscribed, the $200 plan was sold as Max at twenty times Pro’s usage, with a five-hour window and a weekly one. The page now describes it as ten times Pro’s weekly usage with no five-hour limit for verified accounts, and the name Max belongs to the $100 tier alone. Nothing in the price changed. What changed is the shape of the allowance, and my account’s meter shows the day it happened.

Four mechanics decide what those rows mean in practice.

The five-hour window. On Pro and Max, usage is metered inside a rolling five-hour session as well as across the week. This is the limit a single busy afternoon hits: it fills fast and it resets fast. On Ultra it is gone, and on my account it vanished mid-study.

The weekly window. Every paid tier has one. On my account it reset every Wednesday at the same local hour, 01:00Z through October and 00:00Z after the US clocks changed on November 1, and a second weekly window tracks it reading for reading; the vendor has renamed that second window twice since September, so I call it only the second weekly window. The weekly window is the one that caps a sustained week, and the one the exhibits below are about.

One pool across surfaces. The allowance is the account’s, not Claude Code’s. The terminal, the desktop app, the browser and any unattended worker logged in as you all draw from the same windows, which is why my “steering” weeks include the workers that were alive beside me.

Past the cap. When a window fills, the tier stops until the reset unless you buy usage credits, which Max and Ultra can; both also include a monthly API credit, $100 and $200, that pays for overflow through the API at list price rather than through the plan.

How the windows behaved on the $200 tier#

Eight completed weekly windows and one open one. The observed peaks were 80, 46, 30, 100, 71, 51, 48 and 55%, and the open week stood at 20% after a day and a half. The 80 is a window first sampled when it was already at 52%; the 100 is the week of October 7, when the governor’s telemetry had stopped on the 10th at 15% and the transcripts record the cap being hit in interactive sessions on the 12th and 13th. That is the only weekly cap the account produced in the period. The August caps were earlier and different: a five-hour alert at 98% on August 22, a weekly alert at 95% on August 25 and a model-scoped weekly alert at 97% on August 30, all during heavy hand-driven use before any retained series began.

Two rates fall out of these windows and they are the rates the picker uses. The busiest pace any completed window sustained for three hours was 2.3 points of the weekly allowance an hour, in the week of October 28 with two workers alive, so a busy hour at the keyboard costs up to 2.3% of an Ultra week and one Ultra allowance is about 43 such hours. Over the four governor-era windows the account consumed 210 points across 837 alive worker-hours, a quarter of a point per worker-hour, so an always-on worker costs about 42% of an Ultra week; two fit, four do not. Both rates include my own typing, because the account is one meter, so both are upper bounds per worker and per hour.

The five-hour window is the other half of the story, and on this account it has an ending.

Seventeen windows in the last week it existed, median 4%, the worst 23% from one busy afternoon. At one or two workers the five-hour window was never the constraint. Its last reading was at 02:28Z on October 31. The same state file carried weekly readings for another twelve hours with no five-hour window beside them, and no reading since, through November 12, has had one. The governor drops a window the moment the vendor stops reporting it, so there is no later file that could hold one. The account that was Max 20x, with three windows, now has the Ultra shape the pricing page describes, with the weekly windows alone. I did not change the subscription; the tier changed under it.

What the weeks held, in human terms, is the last exhibit. Fourteen ISO weeks from late July: how many terminal sessions on Anthropic’s models ran an hour or longer, what the CLI estimated the week’s sessions would have cost at list price, and what the weekly window reached.

The line does not follow the bars tidily, and that is the point: the one or two workers alive all week, and which models the sessions used, matter as much as the hours. The dollar column is what one plan-dollar bought. Ultra costs $46 a week; in the weeks with more than $500 of estimated use the account drew between 20 and 94 times that in list-price inference, and the quietest week over $100 still drew four times. Per point of the weekly allowance the estimate ranged from $8 in October to $33 in late September, a four-fold spread that comes down to which models were in the mix, and the picker shows a range for that reason rather than a number.

What Pro and Max would have done with the same weeks#

Take the vendor at its word. Ultra is ten units a week, Max five, Pro one, so a week that used 51% of my allowance would have used 102% of a Max week and 510% of a Pro week. Capping the modelled rows at 100 and spreading each week’s use evenly across its 168 hours gives the hour the cap would have landed.

On Max, five of the eight completed weeks cap. The 71% week caps late on day five, the 51% and 48% weeks are a coin toss either side of the line, and only the 30% and 46% weeks, the quietest of the period, fit with room. The five-hour window, with a quarter of the allowance my account had, would not have bound once in the seventeen windows sampled.

On Pro, all eight weeks cap, most of them by Thursday. The 30% week, my lightest, reaches a Pro cap at hour 56, and the 71% week at hour 24. The five-hour window would have bound in six of seventeen windows at one or two workers, and the 23% afternoon would have been a stop at 100%. Pro’s limits are the limits of a tool you use in sessions, not of one you leave running.

These are modelled rows. If the real Pro allowance is more generous than a tenth of Ultra’s the cap hours move later; if the model-scoped windows bite differently at the smaller tiers they move earlier. Treat the shape as the finding, the hour as the estimate.

Where Pro is the right plan, and what its limits are#

Pro is right when the agent is a tool you pick up, not a worker you leave running: a few evenings a week, a problem at a time, with a human reading every diff. Pro’s limits are two windows, five hours and a week, and the measured rate puts the weekly one at about four busy hours: five busy hours a week is about 12% of an Ultra week, which is 116% of a Pro week, so at that pace the Pro cap lands around the last day, and anything lighter fits. It is also the right plan to find out whether you like the loop at all, because the five-hour window will tell you within a week whether you are a Pro-shaped user: if it stops you more than once a week, you are not. Annual billing takes it to $17 a month, paid up front.

Where Max is, and whether it is worth it#

Max is right when you work with the agent most days and nothing runs while you sleep. Is Max worth it over Pro? If the five-hour window has stopped you more than once in a week, yes, and the step is the cheapest one on the ladder: five times the allowance for five times the price, plus the API credits. Three of my eight weeks fit it, and those were weeks of 30 to 46% on Ultra: a working week with real multi-hour sessions and one worker alive. One always-on worker alone is about 84% of a Max week, which is tight but holds; a worker and a working week of steering does not. The five-hour window on Max never bound in anything I measured, so Max’s practical limit is the weekly one. Max also includes $100 of API credits a month, which is a quiet way of saying the overflow is already paid for at list price. If you would pay $200 for Ultra mainly for the headroom, try Max for a month and read the meter on Wednesday morning.

Where Ultra is#

Ultra is right when anything runs unattended, or when you steer for thirty hours a week, or when a cap mid-task costs you more than $100 a month would. On my account it held every week at 30 to 80% with one or two workers alive and a person typing, and it capped once in eight weeks, in a week the governor was not watching. The structural difference is the one the pricing page footnotes: no five-hour window. For a single developer that means the busy afternoon is never the problem; for anyone running an agent through a backlog it means the only meter to watch is the weekly one, and it resets on a fixed day. Two always-on workers is where Ultra gets tight, and four is over, so Ultra is one developer and a small fleet, not a large one. Beyond that you want a second account or the API.

When Team seats beat two Max plans#

Team pricing is per person. A Standard seat is $20 a month billed annually or $25 monthly and the page says only “more usage than Pro”; a Premium seat is $100 annual or $125 monthly and five times a Standard seat. Priced annually, a Premium seat costs exactly what Max costs, and buys roughly the same allowance, so for two people who each need a Max-sized week the choice is not about usage at all. Team wins when you want one bill, central administration and shared conventions; two Max plans win when each person wants their own API credits, which the page lists on Max and Ultra and not on Team seats, or when one of the two actually needs Ultra, which Team does not sell. The Standard seat is the interesting one: it is Pro’s price with “more usage than Pro”, so for a team of occasional users it is the cheapest way to put several people on a plan. In the picker the Team rows give every person the same hours and split the always-on workers across the seats; the Standard row is a floor, because the page gives it no multiplier. Enterprise is $20 a seat at twenty seats minimum with usage billed at API rates, which is a different product: predictable per-seat cost and a metered bill, no windows.

When pay-per-token wins#

Below about five busy hours a week with no plan already in hand, pay per token. Five hours of the kind of session I run is $100 to $380 a week at list price by the CLI’s own estimate, which is already past Pro’s monthly price in a week, so the break-even arrives early; but a lighter user, one who asks and reads rather than steers, can spend a month under $20. The API has no windows and no five-hour clock, which also makes it the right answer for anything that must not pause: a scheduled job, a hook, a review bot. For everything else the plans are cheaper by a wide margin. Ultra’s $46 a week bought 20 to 94 times that in list-price inference in my busy weeks, and even the quiet weeks cleared the plan price. The plan finder puts these tiers next to the other vendors’ plans on the same tokens-per-dollar axis if you want the wider comparison.

Tested with#

Claude Code 2.1.322 and Codex CLI 0.167.0 were the installed versions when this was written. The usage telemetry runs from September 16 to November 12, 2026, on one $200 account; the transcript tables run from July 9 to October 31 and come from the earlier Codex study. The plan names, prices and multipliers are claude.com/pricing as it read on 2026-11-11, on two reads the same day; the page changes without a dated changelog, and the Max 20x to Ultra rename is the proof. The five-hour window’s last reading on the account was 2026-10-31 02:28Z, and the latest reading, 2026-11-12, reports weekly windows only. A weekly check compares the installed versions and the age of the telemetry against this page and flags it for a refresh when either tool moves by a minor version or the data passes ninety days. Every number in this guide comes from the tables under the charts above, which are the account’s own readings; if you want the full set of readings behind them, ask me through the contact page.

— Jed