Anthropic just told Claude Code users that weekly usage limits are permanently going up by 25 percent starting September 14. On the surface this looks like a gift. In practice it is a cut.
Here is why. Back in May, Anthropic gave Claude Code, its AI coding assistant, a temporary 50 percent boost above the normal weekly limit, meant to be short lived. That boost got extended several times over the summer. Now Anthropic is finally ending it, but instead of removing it and restoring the old limit, it is swapping it for a smaller, permanent 25 percent bump. Do the math and anyone using Claude Code today ends up with about 17 percent less usage available after September 14 than they have right now. Anthropic's own account admitted as much once users worked out the numbers themselves.
This matters because a growing number of companies rely on Claude Code every day. The tool lets developers describe what they want in plain language and have the AI write, test, and fix code, and businesses have leaned on it harder than almost any other AI product to speed up software work.
The Claude Code pullback is not a one-off. Days earlier, OpenAI brought back a strict five-hour cap on its own coding tool, Codex, for regular ChatGPT Plus subscribers, ending a six-week stretch where only a looser weekly limit applied. Two of the largest AI companies in the world tightened the same kind of product within days of each other.
The reason is money. Independent research from SemiAnalysis found that a $200-a-month Claude subscription can deliver computing power worth roughly $8,000 if bought at standard rates. A ChatGPT plan at the same price can be worth even more, close to $14,000. Both companies have been losing money on their heaviest users for a long time, and coding tools attract exactly that kind of constant, all-day use. Limits are how they slow the bleeding without officially raising the sticker price.
Businesses that assumed a flat monthly fee meant predictable costs have already been surprised elsewhere. Uber reportedly burned through its entire year's AI coding budget in about four months after rolling Claude Code out to thousands of engineers. Microsoft reportedly pulled back its own internal access after token-based costs spiked to between $500 and $2,000 per engineer each month, pushing staff toward its own GitHub Copilot tool instead.
The lesson for any manager weighing AI coding tools is simple. A subscription price is not a cost ceiling. Whatever a plan advertises today can move, and often already has moved several times this year, sometimes up and sometimes down depending on how much the AI company can absorb. Anyone building a workflow or a budget around Claude Code, Codex, or similar tools should track actual usage rather than trust the sticker price, because these companies are still searching for a pricing model that does not lose money on their best customers.