# OpenAI compatibility

> Understand the supported OpenAI-style surface and the features that remain outside it.

Router supports a focused subset of the OpenAI chat-completions interface. Existing OpenAI clients can connect by changing the base URL and API key.

## Supported today

| Capability                           | Support                                                               |
| ------------------------------------ | --------------------------------------------------------------------- |
| `GET /v1/models`                     | Model edition discovery with Router limits and availability counts.   |
| `POST /v1/chat/completions`          | Text messages with `system`, `user`, and `assistant` roles.           |
| Non-streaming responses              | Assistant text, finish reason, and provider-reported usage.           |
| Streaming responses                  | Ordered content deltas, final finish reason and usage, then `[DONE]`. |
| OpenAI Python and JavaScript clients | Supported through a custom `base_url` or `baseURL`.                   |

## Accepted chat parameters

Router accepts `model`, `messages`, `stream`, `max_tokens`, `temperature`, and the optional Router extensions `provider_selection` and `slippage_tolerance` (integer 0–100, default 30%). Omit `provider_selection` for Auto; use `{offer_id,pricing_revision}` to choose one public computer/model offer. See [choosing a provider](https://router.optimai.network/docs/choosing-a-provider.md). Unknown properties are rejected, which helps catch code written for a broader API surface.

As with the OpenAI API, `stream` defaults to `false`. Long work must stream: a non-streaming request above 512 output tokens or about 24 KB of message text is rejected with `stream_required`, because a long silent response is cut off by network timeouts.

```json
{
  "model": "gemma3:1b",
  "messages": [
    { "role": "system", "content": "Answer briefly." },
    { "role": "user", "content": "What is a neural network?" }
  ],
  "max_tokens": 200,
  "temperature": 0.4,
  "stream": false
}
```

## Outside the supported surface

Router does not currently support tool or function calls, image input, structured output, embeddings, audio, the Responses API, arbitrary message names, or automatic model fallback.

> **Do not assume drop-in parity**
> An OpenAI client is a convenient transport. It does not make every OpenAI endpoint or option available. Keep your integration inside the documented subset.

## Migrate a basic client

```typescript
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: process.env.ROUTER_BASE_URL,
  apiKey: process.env.ROUTER_API_KEY,
});
```

Use an environment base URL ending in `/v1`, then call `client.models.list()` or `client.chat.completions.create(...)`.

## Design for explicit failure

Router preserves your model and explicit provider choice and reports unavailable capacity. Manual selection does not silently fall back to Auto. If a selected price changes, review the new pricing revision before sending a new logical request. If your product wants a fallback model, make that choice in your application, tell the user when it happens, and validate the alternate model’s limits first.

Continue with the [Python](https://router.optimai.network/docs/sdk/python.md), [TypeScript](https://router.optimai.network/docs/sdk/typescript.md), or [cURL](https://router.optimai.network/docs/sdk/curl.md) recipe.