Side by side
| OpenAI | Claude API | |
|---|---|---|
| Limits you set | Monthly hard limit on the organization or a project | Monthly limit on the organization or a workspace |
| Limits set for you | An approved monthly usage limit from your usage tier | A monthly cap per tier: Start $500, Build $1,000, Scale $200,000; the Custom tier has none |
| How the levels combine | Both levels can apply to one request | Workspace limits sit under the organization's; you can't set a spend limit on the default workspace |
| Status at your limit | 429 | 400, with error type invalid_request_error |
| Status at the provider's limit | 429 | 429, with error type rate_limit_error and no retry-after header |
| Error to look for | organization_spend_limit_exceeded, project_spend_limit_exceeded, organization_usage_limit_exceeded | enforced_spend_limit_reached at the tier cap; a message starting "You have reached your specified" at your own limit |
| When access returns | When you raise or remove the limit, or at the next monthly cycle | When you raise or remove your limit; a tier cap lifts at 00:00 UTC on the first of next month, or with a higher limit |
These details were checked against each provider's documentation on the date at the top of this page, and they change. Check the sources below before you build on them.
Alerts on OpenAI don't stop anything
OpenAI separates spend alerts from hard limits. Its documentation is direct: Spend alerts do not enforce a cap.
An alert notifies you and traffic carries on, and alerts remain active alongside a hard limit, which makes them useful as the early warning. If you've only set an alert, nothing is protecting you. More on this in spend alerts vs hard limits.
Expect a small overshoot
OpenAI says enforcement isn't instantaneous, and that recorded spend can slightly exceed the configured amount while the limit takes effect. Treat a provider limit as a strong ceiling rather than an exact one, and set it below the number you really can't go past.
Handle these errors differently from rate limits
This is the detail most likely to bite an agent. A rate limit clears in seconds, so backing off and retrying is right. A spend limit can last for weeks, and retrying does nothing. Anthropic notes that its spend-cap response carries no retry-after header and that retrying, including the SDK's automatic retries, fails until access resumes.
So an agent that treats every 429 the same will sit retrying. Read the error code, and when it's a spend limit, stop the run and tell someone. Note too that on the Claude API your own limit comes back as a 400, so code that only watches for 429 will read it as a bad request.
What these limits don't cover
- They aren't per agent. OpenAI's guide describes organization and project limits, and Anthropic's describes organization and workspace limits. Neither describes a limit on one API key.
- They're monthly. A limit on the month doesn't stop a bad night early in the month. It only ends the damage once the whole month's allowance has gone.
- They stop everything in scope. When an organization limit trips, the healthy agents stop with the broken one.
- Each covers one provider. If you use two, you have two limits and no total.
You can get part of the way with what the providers offer: one OpenAI project or Claude workspace for each agent gives each its own limit. That works for a handful of agents, and it gets awkward with many. Beyond that, teams generally put a gateway in front, compared in LLM gateway budgets compared.
Spend is in early access: you can set budgets and see what each call cost today, and budgets that stop calls are being switched on across the platform.
Questions
- Does OpenAI have a hard spending limit?
- Yes. OpenAI offers monthly hard limits on an organization or a project. When tracked spend reaches the limit, requests return a 429 error. Spend alerts are separate and only notify you.
- What happens when you reach your Claude API spend limit?
- At your tier's monthly cap, requests return HTTP 429 with the error code enforced_spend_limit_reached, and usage pauses until 00:00 UTC on the first of the next month unless you get a higher limit. At a limit you set yourself, requests return HTTP 400.
- Can I set a spend limit per API key on OpenAI or the Claude API?
- Their spend limit documentation describes organization and project limits on OpenAI, and organization and workspace limits on the Claude API, not limits on one key. Giving each agent its own project or workspace is the usual workaround.
- Should my agent retry after a spend limit error?
- No. A spend limit won't clear in seconds, so retrying wastes time. Check the error code, end the run, and tell a person.