What Is an AI Abuse Rate Limit? (2026): A Pattern Cap, Not a Monthly Quota

An AI abuse rate limit is a rule that slows or stops a caller because the pattern looks like abuse, even when that caller has not spent the monthly request budget. A budget counts how much you bought. A retry budget counts how many times a failed call may run again. An abuse limit watches the shape of the traffic: one key across many accounts, a burst that ignores the error, or a prompt that keeps asking for another tenant’s data.

The monthly cap is what an AI request budget is. The repeat of a failed call is what an AI retry budget is. The switch that opens when a tool keeps failing is what an AI tool circuit breaker is.

Claude’s prices below are from claude.com/pricing on October 24, 2026. ChatGPT and Perplexity prices are from directory pages checked October 16, 2026, because chatgpt.com/pricing and perplexity.ai/pro returned HTTP 403 that day. Cursor’s price is from the same October 16 directory record. Published October 24, 2026. We did not measure how often a limit fires.

What the limit is allowed to use

Signal you can name	The limit may	The limit may not
Calls per minute from one key	Slow or stop that key	Spend the rest of a different team’s monthly budget
The same failing prompt, repeated	Stop after the retry budget you wrote	“Try a friendlier wording” and send it again
A request for another tenant’s records	Refuse, and log the key	Answer from the other tenant because the question was specific
A human’s normal workday	Leave it alone if it stays inside the budget and the pattern rules	Treat a long Project as abuse because it is long

flowchart TD
  call["A call arrives"] --> budget{"Inside the request budget?"}
  budget -->|No| stopb["Stop. The month is spent"]
  budget -->|Yes| pattern{"Does the pattern match an abuse rule you wrote?"}
  pattern -->|Yes| stopa["Slow or stop. Log the rule id"]
  pattern -->|No| retry{"Is this a retry, and is the retry budget open?"}
  retry -->|No budget left| stopr["Stop. Do not invent another attempt"]
  retry -->|Yes, or not a retry| run["The call may run"]

A 429 from the vendor is their limit. Your abuse rule is the one you can show an auditor. If you only have the vendor’s message, you do not yet have a rule. Write the pattern in words a second person can check: which key, which window, which action.

Where the rule is written

The long policy. Claude Pro is $17 a month with the annual discount, $200 billed up front, or $20 billed monthly, on the October 24 pricing page. Put the abuse rules in a Project next to an example of a blocked call. The review is Claude review.

One rule, from a paste. ChatGPT Plus is $20 a month on the October 16 directory card. Go is $8 and may include ads. Free can format a single rule. Discard a rule that says “block anything suspicious” and does not name the pattern. The review is ChatGPT review.

The code that enforces it. Cursor Plus is $8 a month, or $80 a year, on the October 16 directory card. The limit belongs in the gateway, in a diff you review. A generated change that deletes the tenant check is a defect. The ladder is Cursor pricing explained.

A search product, which is not your gateway. Perplexity Pro is $17 a month billed annually on the October 16 card. It answers from the web with citations. It does not apply the key rules on your private API. The invoice is Perplexity pricing explained.

A prompt that holds the line:

“Turn the patterns I pasted into abuse rules. Each rule needs a key, a window, and a stop action. Do not add a pattern. If a window is missing, write Window missing and stop.”

What not to delegate

Do not let the model decide that a customer “seems abusive.” That decision is the rule, or a person. Do not raise the monthly budget because the abuse limit fired. Those are different dials. Do not log the full prompt if the prompt holds a customer identifier. Log the key, the rule id, and the time. The strip step is what AI PII redaction is.

FAQ

How is an abuse rate limit different from a request budget?

The request budget is how much you are allowed to spend in a period. The abuse limit can stop a caller who has budget left, because the pattern is one you forbade. A quiet caller can spend the budget. A noisy one can be stopped on the first minute.

How is it different from a retry budget?

A retry budget counts another attempt after a failure you defined, such as a timeout. An abuse limit counts a pattern you defined as not allowed, including a retry loop that ignores the error. When the retry budget is zero, the right action is to stop, not to rephrase the call.

Should the model write the threshold?

It may format a threshold you already chose. A number it picked because “most APIs use 60” is not your limit. Leave the cell blank and write the number from the incident or the vendor’s published cap.

What do you tell the caller?

The rule id and the window, if your policy allows that much. A paragraph about why the system “felt” the call was unsafe is not an explanation. The circuit breaker article is the failure path. This article is the pattern path.

Last verified

October 24, 2026 for Claude. October 16, 2026 for ChatGPT, Cursor, and Perplexity. No separate abuse-limit product fee is recorded in the directory. Published October 24, 2026.

Sources: Claude pricing, ChatGPT pricing, Cursor pricing, Perplexity pricing.

Published by Andrew for AI Tools Optimizer.

Claude is in the chatbots and assistants directory. Directory pages may use affiliate links. A purchase through them may earn us a commission at no extra cost to you.