Reference

Errors

Errors come back in OpenAI’s shape, with a stable code you can match on and a message written for people. A refused or failed call costs nothing.

Checked 26 October 2026

The shape

Response

{
  "error": {
    "type": "rate_limit_error",
    "code": "daily_cap",
    "message": "This call needs up to $0.02 and the key has $0.01 left of its $2.00 daily cap, which resets at 00:00 UTC. …"
  }
}

Match on code, not on message: messages can be reworded, codes stay. type follows OpenAI’s, so SDKs raise the exception you expect.

Every code

Status	Code	When	Fix
401	`missing_key`	No key was sent.	Send it as `Authorization: Bearer sk-nokeep-…`. Make one in the console, or try `sk-nokeep-test` first.
401	`invalid_key`	The key is wrong, or it was revoked.	Copy the key again from where you saved it; the console shows only its last four characters. Make a new one if it is lost.
403	`account_frozen`	Paid features are paused on the account the key belongs to.	Write to us from the contact page.
429	`rate_limited`	The key sent more than 120 requests in a minute (20 for the test key, per address).	Wait the seconds in `Retry-After`, then retry.
429	`daily_cap`	The call could pass the key’s daily cap.	Raise the cap in the console, lower `max_tokens`, or wait for 00:00 UTC (`Retry-After` says how long).
402	`insufficient_credits`	The account’s bought credits can’t pay for this call.	Top up from $5 in the console, or lower `max_tokens`. A plan’s credits stay in the app.
400	`invalid_json`	The body isn’t a JSON object.	Send `Content-Type: application/json` and a JSON object.
413	`too_large`	The request is over 12 MB.	Send pictures as links instead of data URLs, or send fewer at once.
404	`model_not_found`	The model id isn’t one of ours.	`GET /v1/models` lists the ids. Check for a typo or a missing `-uncensored`.
400	`invalid_messages`	`messages` is missing or empty, too long, has an unknown role, or a tool message has no `tool_call_id`.	Send 1 to 2,000 messages, each with a role of system, developer, user, assistant or tool.
400	`unsupported_n`	`n` is more than 1.	Send one request per reply you want.
400	`invalid_reasoning_effort`	The model doesn’t take that `reasoning_effort`.	DeepSeek takes none, low, high or max. GLM takes low or high.
400	`unsupported_tool_choice`	GLM got a named function or "required" as `tool_choice`.	Send auto or none to GLM; it picks its own tools. DeepSeek takes every form.
400	`unsupported_input`	A content part is a kind the model can’t read.	Send pictures to DeepSeek; GLM reads text only.
400	`too_many_images`	More than ten pictures in one request.	Send at most ten.
422	`hard_limit`	The request crosses one of the hard limits in the terms.	Nothing was charged. Drop the message from the history you send back, or the next call stops too.
400	`rejected`	The model couldn’t take the request as sent.	Check the parameters against the model’s list in Models. Nothing was charged.
503	`busy`	The model is busy.	Retry after the seconds in `Retry-After`. Nothing was charged.
503	`model_unavailable`	The model is offline on our side.	Use the other model for now; the status page shows when it is back.
502	`upstream_error`	The model failed before it wrote a word.	Retry in a moment. Nothing was charged.
404	`not_found`	The route doesn’t exist.	The API serves `GET /v1/models`, `GET /v1/models/{id}`, `GET /v1/status` and `POST /v1/chat/completions`.
404	`api_closed`	The API is switched off.	Try again later.

In a stream	Code	When	Fix
—	`stream_cut`	The model stopped before it finished, mid-stream.	What it wrote is billed, nothing more. Retry to get the rest, or continue from what arrived.

Which ones to retry

Retry 429 rate_limited, 503 busy and 502 upstream_error, after the Retry-After seconds when there are any.
Don’t retry the 400s, 401s, 402, 404s or 422: the same request gets the same answer.
A 429 daily_cap waits for 00:00 UTC. Raise the cap instead.

Get an API key Try the free test key: sk-nokeep-test

The shape

Every code

Which ones to retry

What’s on your mind?