Platform

Rate limits

Each API key has a limit on requests per minute (RPM).

Defaults

Tier Default RPM per key
Developer 60
Verified 600
Research 3,000

A key's own limit, set in the dashboard (1 to 100,000), replaces the default. Limits are per key. Two keys in the same project each get their own limit.

How it is counted

  • Windows are fixed calendar minutes (UTC), not a sliding window. The count resets at the start of each minute.
  • The limit counts requests, not tokens.
  • A request counts as soon as it is authenticated, even if it then fails for another reason. Requests rejected with 429 count too.

These count one request each:

Counted Not counted
POST /v1/chat/completions GET /v1/models, GET /v1/models/{id}
POST /v1/responses POST /v1/messages/count_tokens
POST /v1/messages GET /v1/credits, POST /v1/feedback
POST /v1/batches (one per batch, however many requests it holds) Batch requests as they run
POST /v1/workflows/security-review (one per run, however many passes) The passes inside a workflow run
Retrieving, listing, cancelling and deleting; /v1/byo/*

Headers

Inference responses (chat completions, Responses, Messages) carry:

Header Example Meaning
x-ratelimit-limit-requests 600 The key's RPM
x-ratelimit-remaining-requests 599 Requests left in this minute
x-ratelimit-reset-requests 15s Time until the window resets

When you hit the limit

HTTP/1.1 429 Too Many Requests
retry-after: 15
x-ratelimit-limit-requests: 600
x-ratelimit-remaining-requests: 0
x-ratelimit-reset-requests: 15s
{
  "error": {
    "message": "Rate limit reached for this key. Retry after the time in the `retry-after` header.",
    "type": "rate_limit_error",
    "code": "rate_limit_exceeded",
    "param": null,
    "request_id": "req_..."
  }
}

Wait retry-after seconds, then retry. Because windows are fixed, that is the time until the next minute starts.

import time
from openai import OpenAI, RateLimitError

client = OpenAI(base_url="https://api.utmost.site/v1", max_retries=0)

def create_with_retry(**kwargs):
    for attempt in range(5):
        try:
            return client.chat.completions.create(**kwargs)
        except RateLimitError as e:
            time.sleep(float(e.response.headers.get("retry-after", "1")))
    raise RuntimeError("still rate limited")

The Utmost SDKs and the official OpenAI and Anthropic SDKs retry 429 for you by default.

Higher volume

  • Use batches for work that does not need an immediate answer. A batch counts as one request, and its requests are not rate limited.
  • Spread load across keys only if that matches how you want to attribute spend. Each key has its own limit and its own spend counters.
  • Ask for verification or a higher tier if the defaults are too low.