Verified 2026-11-18 · rates follow the live catalog
An LLM gateway sits between your application and the companies that run models. You send every request to one address with one credential, and the gateway decides where it goes, reshapes it if the destination speaks a different dialect, and tallies the cost. Whether that is worth a network hop depends on how many models you use and how much you care when one of them has a bad hour.
Openference serves OpenAI chat completions, the OpenAI Responses API and the Anthropic Messages API on the same host and the same key. When the destination speaks the format you sent, the body passes through unchanged. When it does not, the request is converted, including tool calls and streaming events, and the reply is converted back. See what is and is not translated.
For each model there are multiple upstream routes, across different providers and keys. Selection is weighted and uses cooldowns. A route that returns repeated errors is temporarily taken out of rotation, and healthy routes are preferred. Upstream failures are never charged to your quota.
Every one of the 53 text models in the catalog has a public per-million-token rate, with cached-input prices and off-peak schedules where they exist. The cost-per-request guide works those rates into what a single call costs, and OpenRouter alternatives compares what other gateways charge for the same models and on top of them.
A key can be limited to a subset of models. A request for a model outside that set is refused with a 403, and the model list that key sees contains only what it may call.
Server-sent events are forwarded as they arrive. When a stream has to be converted between formats, the client still gets a well-formed terminal event even if the upstream stream ends abruptly.
| Situation | Direct to the provider | Through a gateway |
|---|---|---|
| One provider, one model | Simpler, one fewer hop, every provider-specific feature available | Little to gain |
| Comparing or mixing several models | A new account, key, bill and client per provider | One key, one base URL, one bill |
| A tool built for a different API format | Needs an adapter you write and maintain | Translation is part of the service |
| Provider-only features | Available as the provider ships them | Only what the gateway passes through |
A gateway cannot make a model better than it is, and it adds latency of its own. If the table's first row describes you, call the provider directly.
Cursor, Cline, Codex CLI, Claude Code, OpenCode, Kilo CLI, Pi, Open WebUI and SillyTavern are listed as supported in the compatibility matrix, with a smoke test for each. Setup guides: Cursor, Claude Code, Codex CLI, OpenCode, Pi, Open WebUI, Cline.
If you already point a tool at an OpenAI-style base URL, the change is one setting. The LLM proxy page walks through exactly what changes in a request.
Model catalog · Pricing · Comparisons · Cost guides · Integration guides · FAQ
Guides: GLM coding plan alternative · Free AI API · LLM gateway · LLM proxy · Cheap AI API · OpenAI-compatible API
Model APIs: GLM API · DeepSeek API · Kimi API · Qwen API · MiniMax API