Skip to content
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog

Claude Max Usage Limits A/B Test: What Is Confirmed

At a glanceQuick answers
What changed on Claude Max?
On some Max 20x accounts SemiAnalysis measured a 19% larger five-hour window and a 20% smaller weekly budget, about 20% less value over a month.
Has Anthropic confirmed an A/B test?
Not publicly. We found no statement on Anthropic’s or Claude’s X accounts or help pages as of October 23, 2026, 01:05 UTC.
Am I affected?
Possibly, if you are a heavy Max user who hits the weekly limit. Light users who only hit five-hour windows may get more room, not less.
Editorial data illustration on a near-white ground: two side-by-side Claude Max 20x meters labeled Account A and Account B, the five-hour bar slightly taller on B and the weekly bar shorter on B, with a large -20% weekly label
Fig 0Same plan, same day, two different weeks. Made by CellCog's image agent, running GPT Image 2.5.

SemiAnalysis, the semiconductor and AI research firm, said on October 22, 2026 that Anthropic appears to be A/B testing usage limits on Claude Max. On two Max 20x accounts measured the same day with the same model, one got a larger five-hour window and a smaller weekly budget, which works out to about 20% less over a month for a heavy user.

The post on X drew about 92,000 views in three and a half hours, and heavy users replied that it matched what they had seen. Anthropic had not commented publicly when we checked.

On this page · 6 sectionsOpen
  1. What SemiAnalysis measured
  2. Why a bigger window can mean less
  3. What is confirmed, and what is not
  4. What to do if you rely on Max
  5. Update log
  6. Sources
Key points5 · 5 min full read
  1. Two identical cards side by side with different bar heights.
    SemiAnalysis measured two Claude Max 20x accounts on the same day with the same model, Opus 5.5, and found different limits on each.
  2. A short bar rising and a tall bar falling.
    Account B had a 19% larger five-hour window (2,076 vs 1,740 million agentic tokens) but a 20% smaller weekly budget (5,266 vs 6,579).
  3. A calendar page with a smaller coin stack beside it.
    Because the weekly budget caps a month, SemiAnalysis puts the monthly API-equivalent value at $9,385 for Account B against $11,726 for Account A.
  4. A speech bubble with a question mark and no reply.
    SemiAnalysis says only some accounts are affected. Anthropic had not commented publicly when we checked its X accounts and help pages.
  5. A document page with a small clock and a calendar icon.
    Anthropic’s help pages already say Max plans have a five-hour session limit and a weekly limit, and that plans are subject to change at its discretion.

§ 01What SemiAnalysis measured

SemiAnalysis’s words: “We detected an A/B test going on that DECREASES your monthly API-value by ~20% while INCREASING your 5h usage limits by ~20%.” It added: “This only affects SOME accounts.”

Animated diagram: two Claude Max 20x accounts start with equal bars, Account B's five-hour bar grows 19 percent while its weekly bar shrinks 20 percent, and its month strip fills only about four fifths
Fig 1Same plan, two accounts: a bigger five-hour window, a smaller week, less in a month
Measure Account A Account B B vs A
Five-hour window, million agentic Opus 5.5 tokens 1,740 2,076 +19%
Weekly budget, million agentic Opus 5.5 tokens 6,579 5,266 -20%
Monthly API-equivalent agentic value $11,726 $9,385 -20%
Multiple of the $200 plan price 59x 47x -12x
Table 1Two Claude Max 20x accounts, same model and day (SemiAnalysis, October 22, 2026)
Weekly budget on two Max 20x accounts, million agentic Opus 5.5 tokensBar chart: Account A weekly budget 6,579 million agentic tokens and Account B 5,266, highlighted, 20% smallerAccount A6579Account B5266Weekly budget on two Max 20x accounts, million agentic Opus 5.5 tokensBar chart: Account A weekly budget 6,579 million agentic tokens and Account B 5,266, highlighted, 20% smallerAccount A6579Account B5266
Fig 2Weekly budget on two Max 20x accounts, million agentic Opus 5.5 tokens

The method matters for reading the numbers. In its October 5 article, SemiAnalysis explains that it reads every usage meter after every request, works out how many tokens of each type fit in each limit, and converts that into an API-equivalent dollar value for one assumed agentic workload. A different mix of input, output and cached tokens would give different dollar figures; the 20% gap between the two weekly budgets does not depend on the mix.

§ 02Why a bigger window can mean less

Anthropic’s usage limits page says paid plans have “a session limit that resets every five hours, and weekly limits that reset once every week.” The five-hour window refills several times a day, so a heavy user runs into the weekly limit long before the windows run out. That is why SemiAnalysis treats the weekly budget as the cap on a month.

For a light user the trade can run the other way. Someone who works in short bursts and never reaches the weekly limit would feel the 19% larger five-hour window and never notice the smaller week. Whether a test like this costs you depends on which limit you hit.

§ 03What is confirmed, and what is not

Claim Status Source
Two Max 20x accounts had different five-hour and weekly limits on the same day Measured SemiAnalysis chart and post
The change affects only some accounts Reported SemiAnalysis post
Anthropic is running an A/B test on Max limits Not confirmed by Anthropic No statement found on its X accounts or help pages
Every Max user lost 20% Not supported Two accounts measured; SemiAnalysis says only some
A higher-priced Max tier is coming Speculation User replies, no source
Table 2Claude Max limits: claim by claim (as of October 23, 2026, 01:05 UTC)

SemiAnalysis has caught a limits test before. In the October 5 article it described one of three identical subscriptions with about 20% lower limits at an unnamed provider, which told it the account was in a very small A/B test and was testing “how to better balance when people hit limits”. The article does not name the provider, so that quote cannot be attributed to Anthropic. SemiAnalysis drew its own conclusion: “It proves providers can silently change subscription limits at any time”.

Anthropic’s Max plan page lists Max 5x at $100 a month and Max 20x at $200, and says: “Price and plans are subject to change at Anthropic’s discretion.” Nothing on that page describes how limits are tested.

§ 04What to do if you rely on Max

  • Track your own meters for a week. Note when you hit the five-hour window and when you hit the weekly limit, and keep screenshots with dates.
  • If the weekly limit is what stops you, the reported test is the one that would cost you, so compare against last month’s pace.
  • For work that cannot stall on a meter, price the same workload on the API, where you pay per token and no weekly budget applies.
  • Ask Anthropic support directly if your limits changed without notice; a public answer from Anthropic would settle the test question.

CellCog’s AI employees run on Claude Opus 5.5 through the API, not a consumer subscription, so a test on Max limits does not change how much work they can do. We have no inside knowledge of Anthropic’s plans.

§ 05Update log

  • October 22, 2026: page opened from SemiAnalysis’s post and chart, its October 5 article, and Anthropic’s help pages.
  • October 22, 2026: explainer scene added.

§ 06Sources

Frequently asked5 questions

Q1What did SemiAnalysis actually measure?

Two Max 20x accounts on the same day with the same model. It reads the usage meters after every request and converts each limit into how many agentic Opus 5.5 tokens fit.

Q2Why does a bigger five-hour window still mean less value?

The five-hour limit resets many times a week, so a month is capped by the weekly budget. A 20% smaller week means about 20% fewer tokens over a month for heavy users.

Q3Is every Max user getting less?

No evidence of that. SemiAnalysis says the change affects only some accounts, and it measured two Max 20x accounts, not the whole user base.

Q4What does Anthropic say Max includes?

Max 5x at $100 a month and Max 20x at $200, with a session limit that resets every five hours and weekly limits that reset once a week. Its help page says prices and plans can change at its discretion.

Q5How can I check my own limits?

Watch the usage meters in Claude’s settings over a week of normal work. Note where you hit the five-hour and the weekly limit, and compare with what you saw last month.

Published 22 October 2026 All Cost, ROI & pricing →