THE DESK

Your AI Plan Shrank And Nobody Sent A Letter

One query took two weeks of his allowance. The bill never moved.

post thisx →linkedin →

A man on a paid plan ran one query. It took two weeks of his allowance. He posted the receipt on Sept 6 and a few hundred people showed up to say the same thing had happened to them.

Nothing about his subscription changed that week. The model underneath it did.

  • "a stark and unexpected amount of usage in a situation that was not new or unique (outside of swapping to Fable 5.1 from Fable 5)" (Reddit, Sept 6)
  • "after 10 seconds I checked the statistics and they showed 99%. I was shocked. Then the research stopped and asked me to pay for more credits or upgrade my plan" (Reddit, Sept 6)
  • "I was burning through my max plan's usage in 20 minutes" (Reddit, Sept 4)
  • "after barely a day usage I am already at 20% of the weekly model cap" (Reddit, Sept 6)

Nobody got a price increase letter. Everybody got less.

The unnumbered plan

An unnumbered plan is one sold as a multiple of a number the seller never prints. We opened the pricing page this morning.

Claude's individual plans as the pricing page showed them on 2026-09-21. The usage column is quoted from that page.
PlanPrice shown todayWhat the page says you get
Free$0Usage limits reset on a rolling five-hour session window
Pro$20 monthly, or $17 a month billed annuallyNo fixed message count stated
MaxFrom $100 a monthChoose 5x or 20x more usage than Pro

claude.com · queried 2026-09-21

  • The paid tiers are sold as five times or twenty times more usage than the tier below
  • The page does not say what the tier below gets. There is no message count anywhere on it
  • Usage is described as depending on how long your conversation is, how hard the job is, and which model you picked

Read that last line again, because it is the whole story. The seller gets to change the model. You do not.

Your gym membership did not go up. They moved the treadmills.

It is not one company

The complaint is not brand loyalty. The same week, the same shape, pointed the other direction:

gpt-6-astra medium however makes my 5-hour limit behave like a fuel gauge on a Hummer (Sept 5)
It needed three 5 hour session limits... It went through the 5 hour session limits in 20 minutes each or so (Sept 5)
Fable 5.1 used 73% of my 5 hour block and they both did the job (Sept 5)

Both did the job. One cost most of the afternoon's allowance.

What our data shows

The newest models carry the worst first-hand reads of the lot, and the files are young. The chips say how young.

  • The parent products both sit in the middle of the pack on deep files. The new model versions underneath them do not
  • Two complaints repeat across nearly every read in this group. It eats the allowance, and it does not follow instructions
  • One person put those together: "Opus 5 keeps doing extra stuff that was never asked for" (Sept 4)

A model that does extra work you did not ask for, on a plan that bills by work done, is not a quirk. It is a meter with an opinion.

Do this Monday

Free, in order. No purchase.

  • Find the usage screen. Every one of these has one buried in settings. Open it, screenshot it today, and write the date on it. That is your before picture
  • Pin your model. If the tool lets you choose, choose, and write down which one you picked. Auto means the seller picks, and the seller is the one being paid by the job
  • Time one real task. Take a job you actually do, run it once, check the usage screen before and after. Now you know what your week holds. Nobody will tell you this number
  • Do the small stuff on the cheap tier. Rewriting an email does not need the expensive model. Most plans let you drop down, and most people never do

What to look out for

  • A new model version is a pricing change with no announcement. The number on your invoice holds. The work it buys moves. Check the usage screen the week after any model you use gets a new version number
  • "5x more usage" is not a quantity. Five times an unstated amount is still unstated. Before upgrading for headroom, measure what you use now, because that is the only number in the sentence you control
  • Auto model selection bills you for the choice. If the tool decides which model runs your job, the seller decides how much of your allowance the job costs. Turn it off where you can
  • The credit upsell arrives exactly when you are mid-job. One read above got the upgrade prompt when the work stopped, not before it started. That is the worst moment to make a purchasing decision and it is not an accident of timing
  • Thin files on new models. These chips sit on young pools, which is why we print the n beside every one. A new model with a bad early file might settle. It also might not. Do not rebuild your week around one until it has a file you can read
  • Watch for the promo that expires into a cut. One read mentions a weekly limit dropping when a promotional rate ended (Sept 5). We have not verified that one and we are not printing it as fact. Check your own account against what you were promised at signup

Where that leaves you

  • Nothing here says these tools are bad. The parent products carry middling files on thousands of reads, which is the honest verdict
  • What moved is the ratio between what you pay and what you get, without a letter, a notice, or a number you could have checked in advance

Screenshot your usage page today. That is the receipt you will want the next time a plan quietly gets smaller.

If somebody you know is about to upgrade for more headroom, send them this first. Ask them what their current number is.

· the Unbought desk · zero affiliate dollars in this opinion

post thisx →linkedin →