Anthropic is testing different Claude limits on identical plans: a gap of up to 20%

Anthropic is running an A/B test on Claude subscription limits, and the results show up in real accounts. According to research firm SemiAnalysis, some users on the same plan at the same price received a five-hour quota about 20% smaller than others. You pay the same and get less work done.

The test came to light not through a company announcement but through an analyst report. In a market where AI assistant subscriptions have become a working tool for developers and analysts, this is unwelcome news.

How the five-hour window works

Claude subscriptions are capped not by the number of messages per calendar day but by rolling five-hour windows. The window opens with your first request, and how much you can get done before the quota runs out depends on context length, the model you choose, and the load on the services. Anthropic does not state the exact limit in tokens. Max plans are described as multiples of Pro (5x and 20x more), not in absolute figures.

This opacity leaves room for experiments. If the five-hour quota is cut by a fifth, the user will notice only because the limit hits sooner than usual. You can compare your results with someone else's only if your workloads match, and that rarely happens.

Why limits have become a point of tension

A fixed-price subscription runs into simple economics: every request costs GPU compute, and heavy users, especially in agentic coding, can run the model for hours on end. In August 2025, Anthropic had already introduced weekly limits for Claude Code, explaining that a small share of users, under 5%, was running the system around the clock. Since then, limits have remained the company's main lever for balancing demand against available capacity.

A/B tests on that lever make sense from a product standpoint. The company needs to know how much quota it can remove without losing subscribers or triggering a mass exodus to competitors. But testing the price of a product and its volume on paying customers without telling them is a matter of trust.

Key facts from the report:

• data source: SemiAnalysis;
• what is being tested: usage volume within identical paid plans;
• size of the gap: about 20% in the five-hour quota;
• the subscription price is the same for all groups;
• there are no officially published limits in tokens, so the difference is hard to verify.

What subscribers and teams should do

First, start measuring your own usage. Record how long it takes you to hit the limit and on which tasks. If the window starts closing noticeably earlier without any change in your habits, that is a signal. Second, compare notes with colleagues on the same plan with a comparable workload. A gap of a fifth is noticeable even in a rough comparison.

Third, do not build critical workflows on a subscription with no margin. A team that uses Claude as part of its development pipeline would be wise to keep a buffer or move heavy tasks to the API, where the price per token is stated explicitly and does not change depending on test-group assignment. A subscription remains a good value for interactive work, but it gives no guarantees on volume.

Finally, it is worth watching Anthropic's reaction. If the company publicly confirms the experiment and explains its purpose, the test will remain a routine optimization. If not, the question of limit transparency will come back to the company with its next plan updates, this time raised by the users themselves.