We have requested high rate limits because our production system requires that.
For development, it would be nice to be able to give developers a key that has a much lower rate limit so that they can’t accidentally trigger an infinite loop of API calls that costs us thousands of euro.
I know that it is possible to have organisation budget limits which trigger warnings or halt charges at a certain level of organisation API key spend, but teams need more granularity than that. We need to be able to both spend as much as we want at the org level for our prod systems to work while also restricting spend for API keys given to devs.
OpenAI doesn’t take feedback articulated clearly, and removed the obviously useful monthly budget hard limit that used to exist for projects and replaced it with illusion.
One can only think that they deliberately enjoy additional profit from overages, abuse, and lack of management interface, just as they enjoy expiring your credits.
You can use a daemon loop of monitoring the admin endpoint usage by daily buckets for a project, and then have hooks to delete all the API keys for unusual use - as long as that also stays working.
Worth flagging one gotcha if you build that daemon-loop yourself: usage/billing data from the admin endpoints isn’t real-time, there’s a lag (often minutes, sometimes longer under load). So even a well-built “hard cap” watcher can let a burst of requests through between polls - a fast retry loop or a runaway agent can rack up real spend in that window before your hook fires. It’s not a reason not to build it, just means the cap should be treated as “stop it fast” rather than “stop it instantly,” and you probably want the threshold set with that lag in mind rather than cutting it exactly at your real limit. Also worth checking auto-recharge is explicitly off if the goal is a true ceiling - it defaults on for a lot of accounts and will happily refill past whatever cap you thought you set.