# AI errors

> Error codes of the AI API, in OpenAI error format, and what to do for each.

Source: https://docs.pmservers.org/ai/errors/

The AI API returns errors in OpenAI format. OpenAI SDKs read them as normal errors.

```json
{"error": {"message": "The budget for this key or account is spent.", "type": "insufficient_quota", "code": "budget_exceeded"}}
```

## Codes

| Status | `code` | Cause | What to do |
|---|---|---|---|
| `400` | `invalid_request` | The body is not valid. The `message` tells why. | Fix the request. |
| `401` | `invalid_api_key` | The key is wrong, rotated or revoked. | Check the key. |
| `402` | `budget_exceeded` | The key budget or the monthly account budget is spent. | Raise the key budget, or wait for the next month (UTC). |
| `403` | `model_not_allowed` | The key cannot use this model, or the model does not exist. | Add the model to the key, or fix the model name. |
| `413` | `request_too_large` | The request is too large. | Send less input. |
| `429` | `rate_limit_exceeded` | More than 60 requests in one minute on this key. | Wait and retry with backoff. |
| `502` | `upstream_error` | The model service is not available now. | Retry later. |
| `504` | `upstream_timeout` | The model took too long to answer. | Use `"stream": true`. |

## Retry rules

- Retry `429`, `502` and `504` with exponential backoff.
- Do not retry `400`, `401`, `402`, `403` or `413` without a change.