You paid for the month. The cap found you anyway, mid-build, no warning.
On March 31, 2026, Anthropic admitted that Claude Code users were hitting their usage limits "way faster than expected." One person reported burning through the Max 5 quota in about an hour of actual work. Real money out the door every month, and the number you were supposedly buying was invisible until the second you ran into the wall.
That is the whole subscription model, right there. You pay a flat fee for a quantity you can't see, the rules behind it change whenever the vendor feels like it, and you discover the edge in the middle of shipping something.
I want to be clear up front. This isn't Anthropic being evil, or Cursor being greedy. They got cornered by their own pricing structure, and so did their users. The structure is the problem.
A flat fee is a bet, and the house sets the odds
The pitch is comfort. One price, no surprises, use it all month. What you're actually doing is betting your own usage against a cap the vendor controls. And the vendor moves the cap.
Anthropic announced new weekly rate limits on Claude Code on July 28, 2025, live around August 28. Fewer than 5% of subscribers affected, they said, and the stated targets were the people running Claude "continuously in the background, 24/7" plus some account-sharers. Fair enough on paper. Except "fewer than 5%" is a polite way of describing the people who build hard. Lean on a tool, and the flat plan starts to resent you for it.
Spring made it worse. Anthropic trimmed quotas during weekday peak hours, a change one of its own engineers reckoned would touch around 7% of users. A promo that had doubled limits ended on March 28, 2026. On top of that, users reported two bugs breaking the prompt cache, silently inflating their costs by 10 to 20 times and draining quotas faster. Three or four knobs you don't control, all turning at once, and your flat fee quietly buys less every week. Nobody asks you first.
You overpay on the quiet week, you get cut off on the loud one
Here's the bit the marketing leaves out. A subscription only goes your way in two situations: you barely use it, or you hammer it into infinity. Land anywhere in the middle, which is where almost everyone actually lives, and you're bleeding money in one direction or the other.
Slow week. Meetings, a doc, you poke the model twice a day. You paid the exact same as the week you shipped a feature. Gone, for nothing.
Loud week is the cruel one. Deadline's Thursday, you're three branches deep, the model's basically your fourth limb. Guess when the cap turns up. Not the lazy Tuesday. The Thursday. The pricing is built so the tool walks out on you at the precise moment you need it most. Which, if you think about it for two seconds, is the worst conceivable time for a tool to walk out on you.
The Cursor receipt
Cursor went through all of this in public. On June 16, 2025, they swapped the old Pro plan, the one that handed you 500 fast responses then unlimited slower ones, for $20 of usage per month billed at raw API rates. And honestly, the math underneath wasn't insane. Frontier models cost a fortune, and a flat fee can't cover somebody running Claude all day.
The rollout torched the trust anyway. People logged in to find $20 gone in days, sometimes in a single afternoon, on work that felt completely normal. One widely shared report described a team whose $7,000 annual plan was drained in one day of use. On July 4, 2025, CEO Michael Truell put his hand up: "We recognize that we didn't handle this pricing rollout well and we're sorry." Refunds offered.
Most people filed it under "bad communication." That's not the lesson. The lesson is that a flat fee is a fiction draped over usage costs that are very real underneath. The moment the fiction stops covering the actual bill, somebody eats the difference, and it is never the vendor. They cap you, or they meter you and say sorry afterward.
Pay for what you used, see what you paid
The honest version is metered pricing where you can watch the meter tick. Forty cents today because you did forty cents of work. The quiet week barely costs you a thing, because you barely did a thing. The loud week costs more, and that's correct, because the loud week is when the tool actually earned its keep. No quota shrinking at peak hours. No promo quietly dying on March 28. No cache bug quietly tripling your bill.
That's the thing we built Flux to do. One key, routing that right-sizes each prompt to a model that fits it, and you pay for the calls you make. The transparency headers tell you which model answered and what it cost, per request, so the number's never hiding behind some plan tier.
The subscription was always going to find your edge in the end. The only thing it got to pick was whether that happened on a quiet Tuesday while you overpaid, or a loud Thursday when it cut you off cold. You never got a vote either way. That's the part I can't get past.