Messages (Anthropic)

Generate a model response using the Anthropic Messages API shape. This is a second API family alongside Chat Completions, for models that support it. Use it as a drop-in for the Anthropic SDKs and for Claude Code.

POST https://inference.openrelay.inc/v1/messages

Which models support this

Not every model serves the Messages family. Check a model's supported_apis field from GET /v1/models rather than assuming:

{
  "id": "openrelay/glm-5.2",
  "supported_apis": ["chat_completions", "messages"]
}

openrelay/glm-5.2 is the current example: it supports both chat_completions and messages on the same model id. A request to a model that doesn't list messages in supported_apis returns 400 unsupported_api.

Authentication

Send your vl_ API key as x-api-key (the Anthropic SDK convention). The gateway also accepts Authorization: Bearer. See Authentication for both header forms.

curl https://inference.openrelay.inc/v1/messages \
  -H "x-api-key: $OPENRELAY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{ ... }'

Request

The body follows the Anthropic Messages shape. Three fields matter most:

Example request

curl https://inference.openrelay.inc/v1/messages \
  -H "x-api-key: $OPENRELAY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openrelay/glm-5.2",
    "max_tokens": 256,
    "system": "You are a terse assistant.",
    "messages": [
      { "role": "user", "content": "Name three primary colors." }
    ]
  }'

Response

A unary (non-streaming) response is an Anthropic message object:

{
  "id": "msg_…",
  "type": "message",
  "role": "assistant",
  "model": "openrelay/glm-5.2",
  "content": [
    { "type": "text", "text": "Red, green, and blue." }
  ],
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 24,
    "output_tokens": 7,
    "cache_read_input_tokens": 0,
    "cache_creation_input_tokens": 0
  }
}

Streaming

Set "stream": true to receive the response as event-typed Server-Sent Events, matching the Anthropic dialect: each event carries an event: line naming the event type, followed by a data: line:

event: message_start
data: {"type":"message_start","message":{"id":"msg_…","model":"openrelay/glm-5.2","usage":{"input_tokens":24,"cache_read_input_tokens":0}}}

event: content_block_delta
data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Red"}}

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn"},"usage":{"output_tokens":7}}

event: message_stop
data: {"type":"message_stop"}

This is different from the Chat Completions stream, which is data-only frames with no event: line. Key differences:

• event: lines are load-bearing and always present, so the Anthropic SDKs dispatch on them normally.
• Usage arrives in two places: message_start carries input_tokens and the cache counters, message_delta carries output_tokens. Combine both to get the full usage for the request.
• There is no [DONE] sentinel. The stream ends at message_stop.

Usage and cache billing

Usage is reported using Anthropic's field names, which differ from the Chat Completions shape in one important way: input_tokens excludes cache reads. There is no subtraction to do yourself.

{
  "usage": {
    "input_tokens": 28,
    "cache_read_input_tokens": 2496,
    "cache_creation_input_tokens": 0,
    "output_tokens": 83
  }
}

Errors

Errors use the same status codes and error body as the rest of the gateway (see Errors): 401 for a missing or invalid key, 402 for insufficient balance, 403 for an account hold, 404 if the model doesn't exist, and 429/502/503 for transient errors. One Messages-specific case:

Using the Anthropic SDKs

curl https://inference.openrelay.inc/v1/messages \
  -H "x-api-key: $OPENRELAY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openrelay/glm-5.2",
    "max_tokens": 256,
    "messages": [{ "role": "user", "content": "Hello!" }]
  }'

pip install anthropic

from anthropic import Anthropic

client = Anthropic(
    base_url="https://inference.openrelay.inc",
    api_key="vl_your_api_key",
)

resp = client.messages.create(
    model="openrelay/glm-5.2",
    max_tokens=256,
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.content[0].text)

npm install @anthropic-ai/sdk

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic({
  baseURL: 'https://inference.openrelay.inc',
  apiKey: process.env.OPENRELAY_API_KEY, // 'vl_...'
});

const resp = await client.messages.create({
  model: 'openrelay/glm-5.2',
  max_tokens: 256,
  messages: [{ role: 'user', content: 'Hello!' }],
});
console.log(resp.content[0].type === 'text' ? resp.content[0].text : '');

export ANTHROPIC_BASE_URL=https://inference.openrelay.inc
export ANTHROPIC_AUTH_TOKEN=$OPENRELAY_API_KEY
export ANTHROPIC_MODEL=openrelay/glm-5.2

claude            # agentic coding, now on openrelay/glm-5.2

Limits

The request body is capped at 10 MiB; a larger body is rejected with 413 Payload Too Large. See Errors.