OpenAI lets you choose which service tiers each API project can use

OpenAI now lets you choose which service tiers each API project can use, in Platform settings, in an entry its changelog dates 21 October 2026. A service tier is the processing class a single request runs on — OpenAI's error guide names default, flex, priority and fast — and you pick it per request with theservice_tierparameter. Organization settings can now restrict those tiers for every project at once. A request that asks for a tier its project does not allow is refused with a 400 error, so the settings are worth reading before your next deployment rather than after it.

What changed

There are now two levels of control, one inside the other. In Platform settings you choose which service tiers a project may use. Above that, organization settings can restrict service tiers for all projects. The organization has the last word: projects cannot enable tiers blocked by the organization. So if a tier is switched off for your project and you cannot switch it back on there, the block may sit one level up.

The changelog entry points to service tier errors , and OpenAI's error codes guide carries the entry for them, "400 - Invalid service_tier argument". Its cause, in the guide's words: the requested or resolved service tier is not allowed for the project. The API answers with the message "Invalid service_tier argument: The requested service tier is not allowed for this project." as aninvalid_request_error, with error.param set to service_tier.

The guide also says which tiers the restriction covers:

• default, flex and priority— project restrictions apply to these three.
• fast— evaluated as priority. For this check, a request for fast counts as a request for priority, so a project that does not allow priority will refuse fast too.
• Scale Tier— outside this project policy.

These are not the usage tiers — Build, Launch and Grow — that our item on OpenAI's three usage tiers describes. A usage tier is where your organization stands; a service tier is how one request is processed.

What it means for you

A request that names no tier can fail too. This is the line most likely to catch you out: requests that omit service_tier or set it to autocan also return this error if they resolve to a disallowed tier. Withauto, or with no service_tier at all, the request is resolved to a tier for you; if that tier is blocked for your project, you get the same 400 as if you had asked for it by name. Code that never mentions service tiers is not exempt.

Check the settings before you deploy.The guide's first step is to check the allowed service tiers in project settings and in organization settings — both, because the organization can block a tier the project would allow. If your application runs in more than one project, one for testing and one for production for instance, check each of them: the choice is made project by project.

Name the tier you need. The remedy the guide gives is to set service_tierto a tier allowed for the project, or to update the allowed tiers in project settings. For requests that useautoor send no tier, its remedy is to update the project settings so the resolved tier is allowed — a change you may not have the rights to make. A tier you name yourself is one you can compare with the settings in advance.

Handle the error by its parameter, not only its status. A 400 says the request was refused, not why. Read the error body: an invalid_request_error whose error.param is service_tieris this restriction. Treat it as a configuration problem and report it as one. Sending the same request again will not change the settings that refused it.

If you manage the organization, you can now keep each project to the tiers you intend it to use , or restrict tiers for all projects at once. Tell the people who run those projects before you change anything, because their requests will start returning this 400 for any tier you take away.

If you use Ultrafast.The guide's list names default, flex, priority and fast, and leaves Scale Tier outside the policy. It does not mention ultrafast, the tier described inour item on Ultrafast mode for GPT-6.1 Sol, so the page does not tell you whether a project restriction reaches it. Send a test request from each project that uses it before you rely on it.

The source

OpenAI's API changelog, the entry of 21 October 2026, and its error codes guide, the entry "400 - Invalid service_tier argument".