Skip to content

Rate limits ​

Rate limits apply to your organization as a whole, not to each key or project. Splitting work across projects separates the reporting, not the limits.

Limits ​

LimitFreePaid
Requests per minute20120
Concurrent requests24
TTS characters per minute2,0008,000
STT audio seconds per minute6003,600

Your organization moves to Paid once it has bought at least $5 of credit in total. If you need higher limits, contact us and we can raise them for your account.

Each limit is counted separately, and a request must be within all of the limits that apply to it.

Concurrent requests ​

A request counts as concurrent for as long as it is in progress: until a response has finished streaming, or for the whole life of a realtime session. On the Paid tier, any combination of four transcriptions, speech streams and open realtime sessions can run at once.

What each request counts against ​

RequestCounts against
POST /v1/audio/speechRequests, concurrency and characters
POST /v1/audio/transcriptionsRequests, concurrency and audio seconds
WS /v1/realtimeOne request when the session opens, concurrency for its whole life, and audio seconds as audio arrives
GET /v1/models, GET /v1/voices, GET /v1/meNothing

Audio seconds for a file are counted once its length is known, after upload. A 429 at that point means you are within your request rate but over your audio rate.

Rate limit headers ​

Responses from the audio endpoints report the state of the limits they counted against:

http
x-ratelimit-limit-requests: 120
x-ratelimit-remaining-requests: 118
x-ratelimit-reset-requests: 1s
x-ratelimit-limit-seconds: 3600
x-ratelimit-remaining-seconds: 3588
x-ratelimit-reset-seconds: 6m0s
HeaderMeaning
x-ratelimit-limit-<resource>The limit per minute
x-ratelimit-remaining-<resource>How much is left in the current window
x-ratelimit-reset-<resource>How long until it is fully replenished, as a duration such as 20ms, 1.5s or 6m0s

<resource> is requests, characters (text to speech) or seconds (speech to text).

Handling 429 ​

A request over a limit fails with 429 and code rate_limit_exceeded or concurrency_limit_exceeded. The response includes retry-after, in whole seconds. Wait at least that long before retrying. Retrying sooner fails again and counts as another request.

The OpenAI SDKs wait and retry 429 responses automatically. If you handle retries yourself, add jitter, so that many clients do not retry at the same moment.

When a model is busy ​

Each model also has a limit on how much work it serves at once, shared by all customers. If a model is at that limit, the request fails straight away with 503 model_overloaded and a retry-after, rather than waiting in a queue until it times out. These requests are not billed and do not count against your limits. The OpenAI SDKs retry them automatically.

xVoice is a product of Stochastic Inc.