How to Set Usage Limits for an AI Product Without Angering Users
Set usage limits for an AI product by choosing the unit you meter, layering a short window over a longer cap, selling overflow instead of only blocking, and showing the number inside the product. The goal is a limit a user can predict, not a surprise.
Key takeaways
- A usage limit should protect your margin and your uptime, not just stop a user from touching the product.
- Layer a short window over a longer cap, because a single reset time is easy to game and hard to explain.
- Pool tokens across surfaces and sell overflow credits, so a heavy user can keep working instead of hitting a wall.
- Use the token bucket model for per minute rate limits and a monthly spend cap for the worst case.
- Show usage, remaining allowance and the reset time in the product, and return a clear retry-after when a request is refused.
Set usage limits for an AI product by choosing the unit you meter, layering a short window over a longer cap, selling overflow instead of only blocking, and showing the number inside the product. A limit a user can predict builds trust. A limit that appears without warning destroys it.
TL;DR
- A usage limit should protect your margin and uptime, not just stop a user touching the product.
- Layer a short window over a longer cap, because one reset time is easy to game and hard to explain.
- Pool tokens across surfaces and sell overflow credits, so a heavy user keeps working instead of hitting a wall.
- Use a token bucket for per minute limits and a monthly spend cap for the worst case.
Disclosure: ZeroRaised publishes this guide and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 2 Nov 2026; ZeroRaised has no commercial relationship with the companies named.
Usage limit is a rule that caps how much of a metered resource, such as requests, tokens or money, a customer can consume in a period.
Why does an AI product need usage limits?#
Every AI product has three exposures that grow with use: the cost of the model call, the risk that one account abuses the service, and the fairness problem when a few users consume a shared resource.
Anthropic lists two kinds of limit on its API. Spend limits cap monthly cost, and rate limits cap requests over a period, set by usage tier and rising automatically with history (Anthropic). New organisations may start in an Evaluation tier with lower limits, which Anthropic describes as "part of how Anthropic prevents fraud and abuse".
The spend cap answers "what is the worst month?", and the rate limit answers "what happens if someone loops a script?". A limit that answers only one leaves half the exposure open. Anthropic also keeps a temporary limit on newer Claude Ultra accounts until it verifies them (The Next Web).
What limit patterns actually work?#
The useful patterns stack.
Rolling window plus a weekly cap. Claude uses a rolling five-hour session window, and paid plans add weekly limits on top. Activity across web, desktop, mobile and Claude Code draws from the same pool, and the FAQ says "there's no fixed message count" (Claude pricing). The short window smooths bursts; the weekly cap catches the user who never stops.
A pooled token allowance. From December, OpenAI's Plus, Pro and Business allowance is shared across ChatGPT and Codex, measured in tokens. Advanced models use more, and instant chat does not draw from it (OpenAI Help Center).
Overflow credits. At the limit, a Claude user can wait, move up a plan, or turn on usage credits at standard API rates (Claude pricing). OpenAI offers waiting, upgrading or buying credits. Credits turn a hard stop into a decision.
A spend cap. Anthropic's monthly caps are $500 on Start, $1,000 on Build and $200,000 on Scale, and users can set a lower limit (Anthropic). Give the buyer a ceiling they control.
Tiered limits for new accounts. A lower starting limit, raised after a payment history or verification, keeps abuse expensive without punishing a real customer.
Clear headers and a usage page. Anthropic's headers show the limit, the amount remaining and the reset time, and it returns 429 errors, usually with a retry-after header telling the client how long to wait (Anthropic).
How does the token bucket algorithm work?#
The token bucket enforces a steady rate. Capacity is continuously replenished up to a maximum rather than reset at fixed intervals. A limit of 60 requests per minute might be enforced as one request per second, so short bursts can still trigger errors (Anthropic).
Anthropic measures limits in requests per minute, input tokens per minute and output tokens per minute. For most models, cached input tokens do not count toward the input limit, so 2 million input tokens per minute with an 80% cache hit rate allows 10 million total input tokens per minute (Anthropic). If your product uses caching, model that headroom before you set the cap.
How should you meter and bill usage?#
A limit is only as good as the number behind it.
Stripe describes usage-based billing as charging customers based on usage. Metronome is now part of Stripe and is its primary usage-based billing platform, recommended for new integrations, including prepaid credits, credit burndown and real-time usage visibility. Basic usage-based billing on the Meters API remains supported, with resources to record usage, offer billing credits and set alerts when a customer exceeds a threshold (Stripe).
Meter the same unit you enforce, so the invoice and the in-product counter cannot disagree. And alert before the hard stop: a warning at 80% is worth more than a refusal at 100%.
If you sell credits, be explicit. OpenAI says credits pay for eligible usage after included limits are reached and do not transfer between accounts (OpenAI Help Center).
How do you roll limits out without angering users?#
- Pick the unit that matches cost. Tokens when prompt length drives the bill, requests when each call has a fixed cost, tasks when work is variable.
- Set a default ceiling from real data.
- Layer a short window over a longer cap. Five hours plus a weekly cap catches bursts and slow burns.
- Add overflow credits so a user can pay to continue instead of stopping dead.
- Show usage, remaining allowance and reset time where the work happens, not only in billing.
- Fail gracefully. Return a clear error with a retry time, and never lose work because a counter ran out.
- Start new accounts lower and raise automatically with payment history or verification.
- Grandfather existing subscriptions where you can. OpenAI lets subscriptions bought before its card and device restrictions continue under an exception (OpenAI Help Center).
What this means for bootstrapped founders#
You cannot absorb an unbounded bill. One scripted loop against a generous free tier can erase a month of margin, and a surprise invoice is harder to recover from than a visible ceiling.
Start small. A conservative default, a clear upgrade path and a usage page are enough early. Add per feature limits only when one action costs much more than the rest.
The guide to pricing an AI product covers how the limit becomes packaging, and the LLM cost guide covers cost. If you are choosing a provider, read the LLM API comparison. The Claude and OpenAI plan analysis shows how the big labs structure the same trade-off.
Where ZeroRaised fits#
ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you sell an AI product, submit it here and state your limits on the page, because a clear ceiling is a selling point. The newsletter carries practical analysis like this. We make no promise of traffic or results.
Sources#
Frequently asked questions
What is a usage limit for an AI product?
A usage limit is a rule that caps how much of a metered resource, such as requests, tokens or money, a customer can consume in a period. Good limits combine a short rate window, a longer allowance and a spend ceiling, so cost and abuse are bounded without surprising a normal user.
Should I limit requests or tokens?
Limit whatever maps most closely to your cost. If long prompts and long outputs dominate the bill, tokens are the honest unit. If each call has a fixed backend cost, requests or tasks are simpler. Many products need both: requests for burst control and tokens or credits for the monthly allowance.
What is the token bucket algorithm?
The token bucket algorithm keeps a capacity that is continuously replenished up to a maximum, rather than resetting at fixed intervals. A limit of 60 requests per minute might be enforced as one request per second, so short bursts can still trigger rate limit errors.
Do usage limits reduce revenue?
They can, if the limit is too tight or the user cannot raise it when they need to. The safer design is a visible default ceiling, an easy way to buy more, and a clear message before the limit is reached, so the limit prevents surprises rather than blocking adoption.
How do I bill usage limits with Stripe?
Stripe supports usage-based billing, and its documentation now recommends Metronome as the starting point for new integrations, including prepaid credits, credit burndown and real-time usage visibility. Basic usage-based billing on the Meters API remains supported, with resources to record usage, offer billing credits and set alerts when a customer crosses a threshold.