# Error reference

> Every error code the platform returns.

All errors share one envelope (`error.message`, `error.type`, `error.code`). Branch on `code` — it stays stable where messages don't.

## Your request

| Status | Code | Where |
| --- | --- | --- |
| `400` | `bad_request` | Malformed input — the message says exactly what to fix. |
| `400` | `model_not_supported_for_endpoint` | Right model, wrong endpoint — an embedding model sent to `/v1/chat/completions`, or the reverse. |
| `400` | `batch_not_supported` | This model can't be used with [batch](/docs/batch). Send the requests to `/v1/chat/completions` instead. |
| `413` | `request_too_large` | Inference body over 10MB; audio uploads over 50MB. |
| `415` | `unsupported_media_type` | The body was sent with a content type we don't read. Send JSON, with `Content-Type: application/json`. |
| `422` | `guardrail_blocked` | Account guardrails stopped sensitive data leaving. |

## Your account

| Status | Code | Where |
| --- | --- | --- |
| `401` | `invalid_api_key` | Missing, revoked, disabled or expired key. |
| `402` | `budget_exceeded` | Hard spend cap reached for the period. |
| `403` | `forbidden` | Key lacks the required scope, or the account is suspended. |
| `403` | `model_unavailable_under_budget` | A hard cap is armed and this model can't be metered, so serving it would bypass the cap. |
| `429` | `rate_limit_exceeded` | Requests/min, tokens/min or concurrency cap. Has `retry-after`. |
| `429` | `rate_limited` | Too many failed authentication attempts from one source — check the key you're sending. One budget covers the gateway, the management API and MCP, so a bad key spent at one shows up at all three. Has `retry-after`. |

## The model

| Status | Code | Where |
| --- | --- | --- |
| `404` | `model_not_found` | No such model, or not one your key can use. |
| `404` | `no_regional_route` | Your account keeps traffic in-region and this model has no route there. |
| `503` | `no_live_deployment` | The model is in the catalogue but nothing can serve it right now. Safe to retry. |

## Us

| Status | Code | Where |
| --- | --- | --- |
| `404` | `not_found` | No such resource *for you* — Roar never reveals whether it exists for someone else. |
| `502` | `upstream_error` | The model could not be reached, including after failover. Safe to retry. |
| `503` | `service_unavailable` | A problem on our side, not with your request or your key. Has `retry-after`. |
| `504` | `upstream_timeout` | The model took too long; safe to retry. |

<Note>
A `503 service_unavailable` never means your credentials are wrong. If we can't reach our own control plane we say so, rather than returning `401` and sending you off to rotate a key that was fine.
</Note>

Include the request id when contacting support. Every response carries it twice, under `x-request-id` and `x-roar-request-id` — same value, so either will do. If you use an official OpenAI SDK it is already on the error object as `error.requestID`.