API overview

The gateway speaks the native wire formats below. Use whichever your existing client already supports — requests and responses pass through byte-for-byte.

Wire formats & base URLs

Format	Endpoint	Base URL
Anthropic Messages	/v1/messages	https://app.inferencehub.tech
OpenAI Chat Completions	/v1/chat/completions	https://app.inferencehub.tech/v1
OpenAI Responses	/v1/responses	https://app.inferencehub.tech/v1
OpenAI Images	/v1/images/generations	https://app.inferencehub.tech/v1
OpenAI Embeddings	/v1/embeddings	https://app.inferencehub.tech/v1
Rerank	/v1/rerank	https://app.inferencehub.tech/v1

The one rule that prevents most 404s: the Anthropic Messages wire uses the bare origin as its base URL (the client appends /v1/messages itself), while the OpenAI-format wires use the /v1 base (the client appends /chat/completions, /responses, …).

Authentication

One key — the same sk-prov-live-… value — wherever your client expects its provider credential:

OpenAI-format clients: as the OpenAI API key, sent as an Authorization: Bearer header.
Anthropic-format clients: as the Anthropic API key / auth token — the x-api-key header or ANTHROPIC_AUTH_TOKEN value.

Keys are scoped and revocable — rotate freely from the portal, and revoke immediately if a key is ever exposed.

Discover your models

The gateway serves whatever models are enabled on your plan. List the live ids — the .data[].id shape is common to the Anthropic and OpenAI /v1/models responses, so this works for any client:

curl -s https://app.inferencehub.tech/v1/models \
  -H "Authorization: Bearer sk-prov-live-YOUR_KEY" | jq -r '.data[].id'

Streaming

All three text surfaces stream with "stream": true, exactly as the native APIs do. During long quiet stretches the gateway may insert SSE comment lines to keep the connection alive — standard SSE clients (and the official SDKs) ignore comment frames automatically, so no handling is needed.

Byte-faithful pass-through

Your request and the model's response pass through unchanged on the native API — no prompt injection, no quality degradation, and the response's model field names the exact model that served it. Provider-side prompt caching passes through too: Anthropic cache_control breakpoints earn the cache read discount on the Messages wire, and the GPT models' automatic prefix caching applies on the Responses wire. One documented exception to pass-through: open-weight models used inside Claude Code are served through a format bridge (the provider only offers them on the OpenAI-style API), which converts the envelope both ways — the model's output and your billing are untouched.

For parameter-level reference, use the upstream docs for each format — every documented field passes through: Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses. Model-specific limits live with the model — see Models.