Your AI Tools Can Spend Your Money While You Sleep. Here’s Who Lets You Stop Them.
TL;DR: Simon Willison argued this week that pay-by-usage services need hard budget caps, not warnings — because “soft caps will not cut it.” So we checked the AI spending limits on the tools our readers actually touch. Anthropic enforces a monthly spend cap by default and lets you set a lower one. OpenAI has hard caps, but you have to switch enforcement on. AWS started enforcing them on 16 September, for some new customers only. If you’re on a subscription you’re already protected. If you’ve added an API key, you may not be.
The argument for hard AI spending limits
Simon Willison’s post of 3 October is short and the point is blunt: usage-billed services should cut you off at a number you set, and “soft caps … will not cut it.”
His description of the failure is the one worth remembering, because it’s the shape every runaway bill takes:
“wake up to an email sent at midnight … find that, while they slept, their rogue service had consumed several hundred (or several thousand) more dollars.”
An alert is not a limit. An alert tells you something happened. A cap stops it happening. That is the whole distinction that makes AI spending limits useful or useless. Willison’s proposed default is a cap that’s on unless you tick a box reading “Remove the budget cap. My application will not be shut down if I exceed the configured budget limit.”
That matters more than it used to, because the thing spending the money now runs on its own. An agent in a loop at 2am doesn’t stop to wonder whether this is getting expensive.
Who actually enforces AI spending limits
We checked the documentation rather than the marketing. Here’s where AI spending limits actually stand on the tools our readers touch.
| Service | Hard cap? | On by default? |
|---|---|---|
| Anthropic API | Yes | Yes — a monthly cap per tier, plus you can set a lower one |
| OpenAI API | Yes | No — you must switch enforcement on |
| AWS | Yes | Partly — since 16 September, some new customers |
| Subscriptions (ChatGPT Plus/Pro, Claude Pro, Lovable, Replit) | n/a | Inherently capped — see below |
Anthropic: a real cap, on by default
Anthropic’s documentation separates the two things most people conflate when they talk about AI spending limits: “Spend limits set a maximum monthly cost an organization can incur for API usage”, and rate limits, which are about speed.
Every tier below Custom carries a monthly spend cap — $500 on Start, $1,000 on Build, $200,000 on Scale. And the behaviour at the cap is unambiguous: “Once you reach your tier’s spend cap, API usage pauses until 00:00 UTC on the first day of the next month.”
Requests then return HTTP 429 with the error code `enforced_spend_limit_reached`. One detail is quietly important: the response carries no `retry-after` header, and Anthropic’s docs spell out why that matters — “Retrying, including the SDK’s automatic retries, fails until access resumes.” A runaway script cannot accidentally retry its way past the cap. That is what AI spending limits look like when someone has thought about them properly.
You can also set your own lower limit on the Billing page. Hit that one and you get an HTTP 400 beginning “You have reached your specified API usage limits”.
OpenAI: hard caps, but opt-in
OpenAI supports monthly spend limits at organisation and project level — and crucially, a spend limit “can monitor spend without enforcement, or you can enforce it as a hard limit so API requests fail after spend reaches the limit.”
So OpenAI’s AI spending limits exist and they work: enforced limits return a 429 with `organization_spend_limit_exceeded`. But monitoring and enforcing are separate settings, and spend alerts notify you without stopping anything. That is exactly the soft-cap arrangement Willison is arguing against.
If you use the OpenAI API, go and check which one you have. It is the single highest-value two minutes in this article.
AWS: new, real, and not for everyone yet
Per Willison and AWS’s own documentation, AWS now has genuine caps: reach a project’s spend limit and AWS pauses the project, stopping its resources. Paid projects get a monthly limit starting at $20, with notifications at 50%, 75% and 90% along the way.
The caveat on AWS’s AI spending limits is availability. It arrived on 16 September 2026, it’s rolling out gradually, and it covers new customers. If you’ve had an AWS account for years, this probably isn’t you yet.
If you’re on a subscription, relax
AI spending limits mostly don’t apply to you, and it’s worth saying plainly because the headline sounds alarming.
If you pay $20 a month for Lovable, Replit or ChatGPT Plus, the cap is the subscription. You can’t overspend; you can only run out. We’ve written about what running out looks like — ChatGPT Pro’s allowance being halved, Claude’s usage limits, and what Replit’s “Free Mode” actually is — and the common thread is that you hit a wall, not a bill.
The danger starts the moment you do one of two things:
- Add your own API key to a tool, so usage bills to you directly rather than coming out of the subscription.
- Turn on credit auto-reload, so running out quietly becomes topping up.
Both are reasonable things to do. Both move you from “capped by default” to “capped only if you set it”.
What to do in the next ten minutes
- Work out whether AI spending limits are even your problem. Do you have an API key anywhere, or auto-reload switched on? If the answer to both is no, you’re done — go and enjoy your Sunday.
- If you use the OpenAI API, check whether your limit is enforced or just monitored. Monitoring is an alert. It will not stop anything at 2am.
- If you use the Anthropic API, set your own limit below the tier cap. The default cap is real, but $500 is a lot of money to discover by accident.
- Turn off credit auto-reload unless you genuinely need it. Running out is annoying. Topping up silently, repeatedly, overnight, is worse.
- Set the limit to what you’d be cross about, not what you expect to spend. The number’s job is to stop a disaster, not to budget.
Who should care (and who shouldn’t)
- Running an agent on your own API key: this is the article for you. An agent that loops is the exact scenario these caps exist for.
- On subscriptions only: nothing to do. Genuinely. Come back when you add an API key.
- Running something that must not go down: a hard cap is a kill switch, and a kill switch can fire at the worst moment. That’s the real trade-off, and it’s why these things are configurable.
- Teaching yourself with free tiers: set AI spending limits anyway, before you need them. It’s a two-minute habit that costs nothing while you’re small.
- Building for a client or employer: find out whose card is on the account before you start. The answer changes how carefully you want that cap set.
Our take
We think Willison is right, and the reason is a change in who’s spending.
When a human typed every request, soft limits were defensible — a person notices a bill climbing. An agent doesn’t. The whole point of the tools we write about is that they keep working while you do something else, and “keeps working while you do something else” and “stops when it gets expensive” are not the same property unless somebody builds the second one.
Anthropic’s implementation is the one to copy, and the detail that impressed us is the smallest: no `retry-after` header on the spend-cap error, so an automatic retry can’t blunder past it. That is somebody thinking about how this fails rather than how it’s documented.
OpenAI’s AI spending limits are fine and badly defaulted. The feature works; it just isn’t on, and a safety feature you have to find is a safety feature most people won’t have. Willison’s checkbox proposal — cap on by default, opt out explicitly — is the correct shape and nobody has implemented it yet.
The honest caveat: we have not tested these caps by deliberately overspending, and we’re not going to. Everything above comes from the vendors’ own documentation. If you’re relying on a cap for something that matters, set it low and test it on purpose while the stakes are small.
A month ago we watchlisted an unverified claim about an agent running up a five-figure bill. We still can’t verify it. But AI spending limits are worth getting right whether or not that particular story was true — and this week two of the three services we checked had the controls sitting there, unmentioned, waiting to be switched on.
Not sure which vibe coding tool fits what you’re building — or whether you need an API key at all? Take the 60-second Vibe Coding Tool Finder quiz →
FAQ
What are AI spending limits?
They’re monthly caps on what a service can bill you. The important distinction is between a soft cap, which sends an alert while usage continues, and a hard cap, which stops requests when the number is reached. Anthropic enforces a hard monthly cap by default; OpenAI supports one but you have to switch enforcement on.
Can an AI agent really run up a huge bill overnight?
It can if nothing stops it, which is the scenario hard caps exist for. An agent that loops doesn’t notice cost, and an alert sent at midnight doesn’t prevent anything. If you are billing usage to your own API key, an enforced cap is the only thing that actually halts spending.
Do I need AI spending limits on a normal subscription?
No. If you pay a flat monthly fee for ChatGPT, Claude, Lovable or Replit, you hit usage limits rather than a bill. Exposure starts when you add your own API key to a tool or enable credit auto-reload, because both bill you beyond the subscription.
