Error codes and limitations

Errors are returned as JSON with a stable shape:

{ "error": { "message": "Insufficient credits. Top up to continue.", "type": "insufficient_quota" } }

Status codes

Status	`type`	When
`400`	`invalid_request_error`	Malformed JSON, or missing `model` / `messages` / `input`
`401`	`authentication_error`	Missing or invalid API key
`402`	`insufficient_quota`	Credit balance is empty — top up
`404`	`invalid_request_error`	Unknown `model` id (see `GET /v1/models`)
`502`	`upstream_error`	The upstream provider returned an error
`503`	`provider_not_configured`	No API key configured on the gateway for that provider

402 is a pre-check. The gateway verifies a positive balance before calling the provider, so an empty balance fails fast and costs nothing.
502 carries the upstream message. When a provider rejects a request (e.g. its own rate limit or billing), the gateway forwards the provider’s message under error.message with type: upstream_error.
Streaming errors. If an error happens mid-stream, the gateway emits a final SSE event describing it and then closes the stream.
/v1/messages uses Anthropic’s error shape. The Messages surface returns {"type": "error", "error": {...}} with Anthropic error types instead of the table above — not_found_error for an unknown model, api_error for upstream/provider failures — and max_tokens is required there (400 if missing).

The documented endpoints only. The gateway implements Chat Completions, Responses, Messages, Gemini-native generateContent, image generation, embeddings, audio transcription, and text to speech. Other OpenAI / Anthropic endpoints (assistants, batches, fine-tuning, files, etc.) are not implemented, and models do not produce file output.
Audio and video input is modality-gated. Only models that declare the modality accept it — today the Gemini family; see Audio & video input.
No reranking endpoint.
A curated model catalog. See Models & pricing for what’s live; GET /v1/models is the complete list and anything outside it is a 404 — except a bare id on an endpoint that natively speaks one provider (a bare Anthropic id on /v1/messages, a bare Google id on generateContent), which is prefixed with that provider before lookup. You can route any supported model through your own credential — see BYOK.
No rate limiting. The gateway does not impose request or token rate limits; your effective limits are the upstream providers’ (surfaced as 502 upstream_error). Spend is bounded by your credit balance.