Errors

EU GPT returns conventional HTTP status codes and a small JSON envelope. This page lists the codes, the payload shape, and which failures are worth retrying.

Payload shape#

All non-2xx responses return JSON:

{
  "detail": "Human-readable detail string or object",
  "error": {
    "type": "string",
    "code": "string",
    "message": "string"
  }
}

A few 5xx paths bypass the envelope and return a plain text body. Treat the envelope as best-effort: always check the HTTP status, then look for detail or error if present.

Status codes#

400 Bad Request#

The request is structurally invalid. The most common causes:

• input is missing or empty.
• model references an unknown id (when not using auto).
• A UUID field is malformed.

{ "detail": "input is required" }

Do not retry without changes — fix the request first.

401 Unauthorized#

Missing, malformed, or revoked API key. Check:

• The Authorization header is present and starts with Bearer .
• The key is prefixed eugpt_.
• The key has not been revoked in the web UI.
• You are calling the same environment that issued the key (production keys do not work against staging).

Do not retry until you have fixed credentials.

403 Forbidden#

The key is valid but the operation is not permitted. The common cases:

• conversation_id refers to a conversation owned by a different user → returns 404 (we deliberately hide existence).
• project_id refers to a project the user is not a member of.
• The license plan disallows the operation (e.g. trial expired, organisation policy disabled API access).

{
  "detail": "Project membership required",
  "error": "project_forbidden"
}

Retry only after fixing membership or upgrading the plan.

404 Not Found#

The referenced resource does not exist for this user. Both “really gone” and “not yours” return 404, to avoid leaking the existence of other users’ data.

Do not retry — the resource is not coming back.

429 Too Many Requests#

A usage limit was exceeded for the caller. Retry with exponential backoff and a cap; if the response keeps returning 429, slow down and try again later. Contact us if you consistently hit limits during normal operation.

500 Internal Server Error#

An unexpected failure in EU GPT. May be transient.

{ "detail": "Internal error during response generation" }

Retry once or twice with exponential backoff. If it persists, the issue is on our side — check the status page (when available) and reach out.

502 / 503 / 504#

Returned by the load balancer when the backend is overloaded or restarting. Treat them as transient — retry with exponential backoff.

Errors in the stream#

If the failure happens after the response has started streaming, the HTTP status will still be 200 but the stream will terminate with an error event instead of response.completed:

event: message
data: {"type":"error","code":"upstream_timeout","message":"Model timed out"}

Always check the last event of the stream — do not assume success just because the stream opened. Common in-stream error codes:

• upstream_timeout — the model took too long. Retry.
• upstream_error — the model returned an internal error. Retry.
• internal_error — generic platform failure. Retry once.
• cancelled — the request was cancelled (e.g. the upstream conversation was deleted mid-flight).

Idempotency#

The endpoint does not yet honour Idempotency-Key. A retried request generates a fresh response. For interactive use this is rarely a problem; for batch jobs, dedupe on your side using a job identifier and only retry on documented retriable codes.