Verified 2026-11-18 · rates follow the live catalog
Plenty of APIs call themselves OpenAI-compatible. What matters is where that stops. This page lists every endpoint Openference serves, shows the two calls most people start with, and spells out which parameters are carried across to the model and which are dropped.
| Endpoint | Format | Purpose |
|---|---|---|
POST /chat/completions | OpenAI | Chat completions with streaming and tools |
POST /responses | OpenAI | Responses API, used by Codex |
POST /messages | Anthropic | Messages API, used by Claude Code |
POST /embeddings | OpenAI | Embeddings; model availability varies, so check the model list |
POST /images/generations | OpenAI | Image generation, synchronous |
POST /videos | OpenAI | Video generation, asynchronous create |
GET /videos/{video_id} | OpenAI | Poll a video task |
GET /models | OpenAI | Models available to the calling key |
curl https://api.openference.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "GLM-5.3",
"messages": [{"role": "user", "content": "ping"}]}'
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.openference.com/v1",
apiKey: process.env.OPENFERENCE_API_KEY,
});
Every request carries Authorization: Bearer YOUR_API_KEY. One key works for the OpenAI SDKs, curl, and the Anthropic-format and Responses endpoints alike.
GET /v1/models returns the OpenAI list shape, with context_length and max_output_tokens for each model so a client can size its own limits. The catalog currently lists 55 models. If a key is restricted to a subset, the list contains only those, and a request for any other model returns a 403.
Requests to a model that natively speaks the OpenAI format pass through unchanged. When the destination speaks Anthropic's format, the request is converted, and the conversion has edges worth knowing:
response_format has no Anthropic equivalent, so JSON modes are emulated with a system instruction rather than enforced by the model.max_tokens nor max_completion_tokens is set, max_tokens defaults to 4096.frequency_penalty, presence_penalty, seed, logprobs, n and parallel_tool_calls.max_tokens becomes length and tool_use becomes tool_calls.If your code depends on seed for reproducibility or n for multiple choices, test that route before you rely on it.
The Anthropic Messages endpoint serves tools that speak it natively, with Anthropic's own error envelope on failures. The Responses endpoint is what Codex calls; where a model's upstream only supports chat completions, the request is translated and the reply returned as Responses JSON. Setup for each tool is in the integration guides, and the LLM proxy page covers the base-URL differences between them.
Yes. Set base_url (Python) or baseURL (Node) to https://api.openference.com/v1, use your key, and pass an exact catalog model id.
Yes. Responses stream as server-sent events, in the format your client asked for.
No. Models that speak the OpenAI format receive the request as sent. Routes served through Anthropic's format drop the parameters listed above.
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