---
title: Streaming
description: Streaming output across CrossModel's three protocols, and how to handle it on the client.
---

# Streaming

Set `stream` to `true` and CrossModel returns incremental results as Server-Sent Events. The event format differs by protocol, but the client approach is the same: read chunk by chunk, append the text, and handle the end and error events.

## Chat Completions

```bash
curl https://api.crossmodel.ai/v1/chat/completions \
  -H "Authorization: Bearer $CROSSMODEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-v4-pro",
    "messages": [{ "role": "user", "content": "Tell me a short story" }],
    "stream": true
  }'
```

Chat Completions returns `chat.completion.chunk` events. Concatenate `choices[0].delta.content` from each chunk to rebuild the full reply.

## Responses

The Responses streaming endpoint uses event objects with a `type` field, for example:

| Event | Notes |
|------|------|
| `response.created` | The response begins. |
| `response.output_text.delta` | A text increment. |
| `response.function_call_arguments.delta` | A function-arguments increment. |
| `response.completed` | The response ends. |

## Messages

The Anthropic Messages streaming endpoint uses the Anthropic event format, such as `message_start`, `content_block_delta`, `message_delta`, and `message_stop`.

## Error handling

A streaming response can still error after the connection is open. The client must handle error events — don't just wait for `[DONE]` or the end event:

- Chat Completions: you may receive a `data:` chunk with an `error` field.
- Responses: you may receive `event: error`.
- Messages: you may receive an Anthropic `error` event.

<Callout type="warning">
  When a stream fails partway through, don't append the result of an automatic retry onto the output you already have. The safer move is to mark the generation as failed and start a fresh, complete request.
</Callout>