AI Tools

AI Limits & Quotas Reference

Working reference · Wave one

Every AI rate limit, and exactly when it comes back.

Message caps, context windows, API rate limits and reset windows for the tools developers actually hit — each traced to a primary source and dated. The finding that fell out of building it: vendors publish the number when the number is the thing you are buying. Metered products document their limits almost completely — GPU hours, API tokens. Flat subscriptions do not, because there the limit is a cost control rather than the product. Where a limit is undisclosed we say so rather than inventing a figure.

16 tools 85 limits tracked verified 2 Aug 2026

Quota reset forecaster

You hit a wall. When do you get it back?

time to reset
window opened quota returns
The finding

Vendors publish the number when the number is what you buy

Sort the surfaces by how completely their limits are documented and you get a clean gradient, not a muddle. Midjourney sells GPU hours by the hour and documents every plan to the decimal. The APIs sell tokens and publish their tiers in full. Consumer assistants sell a flat monthly subscription — and there the limit stops being the product and becomes a cost control, which is a thing companies describe rather than quantify. Google replaced Gemini’s daily prompt counts with undisclosed “compute” units; Cursor calls its allowance “generous included usage.” Same companies, different disclosure, depending on what sits on the invoice.

Documented The vendor publishes the exact figure on a public page. Quote it and cite it.
Console-only The number exists and is authoritative, but you must log in to your own account to see it. Not quotable in the abstract.
Observed Reported by the vendor in a changelog, blog or support reply, but absent from the reference docs. Reliable, but it can move silently.
Undisclosed The vendor has chosen not to publish a figure. Any hard number you find elsewhere is someone’s guess.
The reference

Every tracked limit

ToolPlanLimit WindowWhen exhaustedEvidence
Methodology

How this page is maintained

  1. Primary sources only. Every figure is taken from the vendor’s own documentation, support centre, changelog or pricing page. Secondary reporting is used to find a claim, never to publish one.
  2. Evidence tier is recorded per limit, not per tool. A single vendor routinely publishes its API limits in full while declining to publish its consumer message caps. Marking the tool as a whole would hide that.
  3. Undisclosed stays undisclosed. Where a vendor has stopped publishing a number, no figure is substituted from a third-party estimate. The gap is the finding.
  4. Every limit carries its own verification date. A figure not re-checked inside 30 days is shown as needing a recheck rather than presented as current.
  5. Superseded figures are kept, not deleted. Readers arrive holding the old number and need to see that it changed.
  6. A tool we could not verify is held out, and named.