SemiAnalysis says some Claude Max accounts get 20% less a week. What’s confirmed

What SemiAnalysis reported

SemiAnalysis, a research firm that tracks the usage limits of AI plans, posted on October 22, 2026 that users reporting lower Claude limits “are not wrong.” In its 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%. This only affects some accounts.”

Its chart compares two Claude Max 20x accounts, measured on the same day with the same model, Opus 5.5:

	Account A	Account B	B against A
Five-hour window	1,740 million tokens	2,076 million tokens	19% more
Weekly limit	6,579 million tokens	5,266 million tokens	20% less
A month at API prices	$11,726	$9,385	20% less

The token counts are for SemiAnalysis’s own mix of agentic work. The dollar figures price a month of that work at Anthropic’s API list prices, 59 and 47 times the plan’s $200 price. SemiAnalysis’s method reads Claude Code’s usage meters after every request and works out how many tokens fill each one. A month depends on the weekly limit, because “the 5-hour limit resets many times within a week, so a month is capped by the weekly limits.”

SemiAnalysis has published figures for Max 20x only, and its post doesn’t say how many accounts it measured.

What Anthropic has said

Nothing about the report. As of October 23, we found no comment on Anthropic’s newsroom, in its help center or from its developer account on X.

About 16 hours before SemiAnalysis’s post, Boris Cherny, who created Claude Code at Anthropic, replied to a user who said Claude Code usage was “draining 5x faster”: “No changes on our end.” He asked the user to run Claude Code’s /usage command and share the output.

Anthropic’s own pages describe the limits like this. Its help center says Max 20x “includes 20 times the Pro plan’s per-session usage allowance,” and Max 5x five times. Max also has “a weekly usage limit that applies across all models,” which resets “at a fixed time each week that is assigned to your account.” Anthropic doesn’t publish its size. The help center and Claude’s pricing page both add: “we may limit your usage in other ways, such as weekly and monthly caps or model and feature usage, at our discretion.”

When Anthropic released Opus 5.5 on September 22, it said: “we’re increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans.”

There is a precedent in SemiAnalysis’s own reporting. Its October 5 article describes one of three identical subscriptions “for a particular provider” that had about 20% lower limits than the other two. The provider, which SemiAnalysis didn’t name, confirmed an “extremely tiny” A/B test on limits, said it “didn’t just decrease limits wholesale,” and said it was testing “how to better balance when people hit limits.”

Not everyone sees a change

Paweł Huryn, who publishes his own measurements of AI plans, disagreed: “From what I see, nothing is being degraded or A/B tested.” He measured Max 20x in early October at $9,060 of API value a month on Opus 5.5, close to SemiAnalysis’s Account B, and wrote that the “decreased” value “is just what I originally measured.” His benchmark isn’t SemiAnalysis’s, so the two sets of numbers aren’t strictly comparable.

Usage can also climb for reasons that have nothing to do with a test. Anthropic’s Claude Code documentation says a session open for hours “can use far more of your plan limits than your activity suggests.” Its reasons include long conversations, which are sent again with every request, cache misses after a break, scheduled tasks, subagents and compaction.

Who it would affect

If SemiAnalysis is right, the test cuts both ways. An account in it gets a bigger five-hour window, so someone who mostly hits the five-hour limit gets more done in a session. Its weekly limit is smaller, so someone who uses Max all week runs out sooner. In SemiAnalysis’s words, “a heavy user gets 20% less from it over a month.”

Pro and Max 5x aren’t in SemiAnalysis’s figures, and it hasn’t said whether they’re part of the test.

How to check your own limits

1. On claude.ai, open Settings > Usage. Anthropic’s help center says it shows when your weekly limit resets next.
2. In Claude Code, run /usage. On a Pro or Max plan, it shows your plan’s usage bars and a breakdown of what used them, such as subagents and MCP servers, with suggested changes.
3. Keep your own baseline. Claude’s meters show how much of each limit you’ve used, not how big the limit is, so they can’t tell you whether your account is in a test. Note how much of the weekly meter a typical day of your work uses, and compare from week to week.
4. If you think your limits are wrong, ask Anthropic. Sign in at claude.ai, select your initials in the lower left, and choose Get help.

If the week runs out early

• Wait for the reset. Claude Code 2.1.234 and later can wait for the limit to reset, then continue the interrupted task.
• Pay for more. On Pro and Max, usage credits in Settings > Usage let you keep working past the limit, with a monthly spend limit if you set one.
• Spend less of it. Anthropic’s documentation says Sonnet “handles most coding tasks well,” and suggests keeping Opus “for complex architectural decisions or multi-step reasoning,” lowering the effort for simpler tasks, and running /clear between unrelated ones.
• Use another account or agent. Enjoy shows the five-hour and weekly limits of every Claude account you’ve signed in to, and when each resets. When one runs out in the middle of a task, Enjoy moves the conversation to another account that still has usage, or waits for the reset and carries on, while your computer is awake with Enjoy open. You can also hand the conversation to Codex or another agent, on that agent’s own plan. On Max, Enjoy keeps how much of each week’s limit you used, and says when Max 5x would have been enough.

Enjoy is free, and it doesn’t change Anthropic’s limits: those belong to your Claude plan.

We’ll update this memo if Anthropic confirms or denies a test, or publishes the size of its weekly limits.