# A 25% Increase That Is Actually a 17% Cut > Anthropic is permanently raising Claude Code weekly limits by 25% on September 14. It is also, by their own arithmetic, cutting them 17% from where they sit today. Both sentences are true, and that is the interesting part. 2026-08-30 · Tim Cadenbach · https://www.tcdev.de/blog/a-25-percent-increase-that-is-a-cut/ --- Sunday morning, and instead of doing literally anything else I'm reading a changelog entry about weekly token allowances. This is what my life is now. The entry says Anthropic is permanently raising standard weekly limits in Claude Code by 25% for Pro, Max, Team and seat-based Enterprise, starting September 14. Good news. Except the temporary 50% boost that's been running since May 13 ends on the same day, so the actual move for anyone using Claude Code today is downward. [Anthropic's own wording](https://www.bleepingcomputer.com/news/artificial-intelligence/anthropic-is-cutting-claude-codes-current-weekly-limits-by-17-percent/): "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." The math is not complicated. Baseline 100, boosted 150, new permanent floor 125. Up against the old baseline. Down against your Tuesday. ## The part I want to be fair about They said the 17% out loud, in their own announcement, unprompted. I want to sit on that for a second because it is genuinely rare. The default corporate move here is a post titled "Increasing Claude Code limits by 25%" with the boost expiry buried in a footnote or, more likely, not mentioned at all. Then you find out in three weeks when your Thursday session dies early and you spend an hour convinced you broke something. Instead they published the subtraction. That is worth more credit than it will get, because the people who notice the honesty are a much smaller group than the people who notice the reduction. Same week, they [cancelled the scheduled Sonnet 5 price increase](https://www.morphllm.com/claude-code-pricing) and made $2/$10 per million tokens permanent instead of moving to $3/$15 on September 1. That is a real, unambiguous, money-in-your-pocket win, and it got maybe a tenth of the attention the limit cut did. Which tells you something about all of us, me included. ## So what's actually wrong with it The mechanism, not the number. The boost landed on May 13. Then it got extended. Then extended again. [Four separate times over the summer](https://www.bleepingcomputer.com/news/artificial-intelligence/anthropic-is-cutting-claude-codes-current-weekly-limits-by-17-percent/), each time pushing the cliff a few weeks out. By the third extension nobody was treating it as temporary anymore. I certainly wasn't. I built a working rhythm at 150, staffed my week around 150, and quietly started assuming 150 was the number. That's the problem with a promotion you renew four times. It stops being a promotion and becomes the baseline in everyone's head, and then restoring an honest permanent number reads as a takeaway. The reference point moved. You can be entirely straight with people and still land in a hole you dug yourself. I don't think this was manipulative. I think it was a capacity decision that got extended each time the numbers allowed, which is a reasonable way to run infrastructure and a terrible way to set expectations. ## The money is not the complaint I ran my own numbers a while back, mostly out of curiosity, and the answer was embarrassing. What I put through Max in a heavy week would cost multiples of the subscription at list API rates. The $200 plan is not where I'm being squeezed. If anything I'm the one being subsidised, and I'd rather that arrangement stayed solvent than watch it get yanked entirely in a year. So no, this isn't a post about wanting cheaper tokens. It's about the denominator. If you build a real working process on top of an agent, whether that's a team habit or an actual automated pipeline, you need to know what your weekly capacity is with enough confidence to plan against it. A ceiling that moves because a promotion lapsed is not a number you can staff against. It's not a number you can put in a proposal. It is barely a number you can explain to a colleague without a diagram. That's the cost here, and it's larger than 17%. ## This is the second squeeze this quarter Worth remembering that the programmatic side already moved. Agent SDK usage, `claude -p`, GitHub Actions, anything non-interactive, [came off the subscription and onto its own metered credit pool billed at standard API rates](https://betterstack.com/community/guides/ai/claude-credits/) in June. Pro gets $20 of credits, Max 5x gets $100, Max 20x gets $200, and once you're through them you're paying list price on top of a subscription you already paid for. I understood the logic then and I understand it now. People were running enormous automated workloads on a flat fee, and that was never going to hold. (I know at least two teams who were, and they knew exactly what they were doing.) But stack the two changes together and the shape is clear. The era where a subscription quietly covered whatever you threw at it is over, and it ended in about a quarter. If your plans assumed otherwise, they need redoing, and there's no version of this where that surprise gets cheaper by waiting. ## What I'm doing before the 14th Measuring. That's it, really. I've never actually written down what my weekly burn looks like, because for most of this year I never came close to the ceiling and had no reason to care. I hit the wall exactly once, on a Thursday in July, in the middle of a refactor I did not want to leave half-finished, and my response was to be annoyed for ten minutes and then forget about it. That was a warning I ignored. So for the next two weeks I'm keeping an actual log, so that on September 14 I know whether I'm in the group where 17% is a rounding error or the group where it's a Wednesday that ends at 3pm. Then I'll move the batch and scripted work off the interactive plan where it belongs, which I should have done in June anyway. If Anthropic ships the "more visibility and control of your usage" they hinted at in the same announcement, half of this becomes a non-issue, because the thing I actually want is not a bigger number. It's knowing where I am against the one I've got.