Hit a cap twice this week on two different assistants, and neither told me how long I had to wait or how many I had left before it happened. One let me back in after what felt like a couple of hours, the other seemed to reset overnight. Has anyone actually measured how these windows behave, or is it deliberately unknowable so you upgrade?
Changing my behaviour helped more than understanding the window. I stopped using the free tier conversationally — no "hi, can you help me with" followed by six clarifications — and started writing one long message with the code, the error and the constraint in it.
Same amount of help, roughly a quarter of the requests. Caps are counted per message, not per word, on every free tier I've used.
Counted per message rather than per word explains a lot about how fast I burn through it. I ask in fragments because that's how I think.
Writing the whole problem out in one go also makes the answers better, which was an unexpected side effect of rationing.
Mostly rolling, in my experience, and that's why it feels arbitrary. A fixed daily reset would be predictable — you'd learn it in two days. A rolling window means your allowance comes back gradually as old requests age out, so the wait depends on when you sent the ones before the cap.
The way to find out for yours is to write down timestamps. Note the time of your first message in a burst, the time you get cut off, and the time the next one succeeds. Three cycles and the shape is obvious. It's tedious and it's the only reliable method, because almost none of them publish the number in a place a user reads.
I paid for one month on two of them specifically to compare, treated it as buying information for the price of two lunches, and cancelled both after.
What I learned was that the free tier of the one I liked less was much closer to its paid version than the free tier of the one I liked more. That's a product decision they've made, and you can't see it from the outside without spending the money. Worth doing once if you're choosing something you'll use daily.
Rolling window, undocumented, changes often. Timestamp three cycles if you must know, batch your questions if you just want more out of it, and never automate against it.
@hotswap_hana · 4mo ago · 3 replies
Worth saying before anyone measures too carefully: these caps change without announcement, sometimes weekly, and often differ by region, by whether you're signed in, and by which model the request lands on.
I kept a note of what I'd measured for one provider and it was wrong within a month — not slightly wrong, half. So measure to learn the shape of the window, don't build a routine around the exact number, and definitely don't build anything automated on top of an undocumented free tier.
Reply
Report
@yaml_yeoman · 4mo ago
That explains why the answers I found from a few months ago all contradicted each other and each was stated confidently.
Reply
Report
@logfile_lena · 4mo ago
Every one of those was probably true when written. Undocumented quota posts have a shelf life of about six weeks.
Reply
Report