Rate limits

Requests per minute

60

Refilled evenly through the minute. Retry-After says how many seconds until the next request fits.

Tokens per minute

200,000

Input and output tokens used in the last 60 seconds. One request can go over once; the next one waits.

Requests at once

32

Requests still running, streams included.

Monthly budget

US$50

Spend in the calendar month, Western Indonesia Time (UTC+7).

• These are the values a new key gets. Whoever creates the key can set lower speed limits and a different monthly budget.
• The same speed limits apply to the organization as a whole, across all its keys. A second key spreads the work but does not add capacity.
• A workspace can have its own requests and tokens per minute and its own budget. They apply on top of the key's.
• To go above these values, write to team@getworg.com.

• 429 rpm_exceeded with Retry-After in seconds. With 2 requests per minute, the third request in a row waited 30 seconds.
• 429 tpm_exceeded with Retry-After: 60. Tokens are counted after each request ends, so one large request can go over the limit once; the next request waits.
• 429 concurrency_exceeded with Retry-After: 1 while too many requests are still running.

The OpenAI SDKs wait for Retry-After and retry on their own. See Retries.

• A key stops at its monthly budget with 402 key_budget_exceeded. A top-up does not help; an admin raises the budget or you use another key.
• A workspace or the organization can have a monthly budget too: 402 workspace_budget_exceeded and 402 org_budget_exceeded.
• A month is the calendar month in Western Indonesia Time (UTC+7).
• When the balance runs out, requests get 402 insufficient_balance until you top up.

Before a request is sent, Worg sets aside the most it can cost: every byte of the request body counted as one input token, each image counted at the most one image can cost on that lane, plus max_tokens as output. Without max_tokens, the output side is the model's output limit or, when that is not known, its whole context window. With n above 1, the output side counts once per choice. Through a combo, the hold is for the most expensive lane in it.

If the hold does not fit in the balance, or in what is left of a budget, next to the requests still running, the request gets a 402 whose message says the amounts:

{
  "error": {
    "message": "This request can cost up to $0.157311 and $0.010000 is available. Set a lower max_tokens or top up.",
    "type": "insufficient_balance",
    "code": "insufficient_balance"
  }
}

The hold is released when the request ends, and only what the request used is charged. The same request with "max_tokens": 100 went through. Set max_tokens to keep holds small, most of all on a lane whose output limit is not published, where the hold covers the whole context window.

If a key spends far more in one hour than it usually does, Worg suspends it to protect your balance. Requests with it then get 403 key_suspended, and the organization's alert email is told.

• The line is at least US$5 in an hour, and more than 10 times the key's usual hourly spend over the past week.
• For a key younger than 7 days, the usual spend is the highest of: the key's own average so far, the average of the organization's other keys, and its monthly budget spread over the month (counted up to US$5 an hour).
• An organization owner or admin reviews the key in API keys. If the spend was expected, resume it. If not, rotate it: a new key with the same limits replaces it.
• A key that spikes again within 24 hours of a resume can only be rotated.