Anthropic Messages API

Make a request

Get a key from Mizan Keys and fund your wallet exactly as in the Quickstart. Auth uses x-api-key — the Anthropic SDK's default header — so a plain baseURL override is all that's needed. Authorization: Bearer mizan_rt_... also works, on this endpoint and on /v1/chat/completions.

One real difference from /v1/chat/completions: max_tokens is required on every request, matching the native Anthropic Messages API.

curl https://api.app-mizan.com/v1/messages \
  -H "x-api-key: mizan_rt_..." \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Say hi in five words." }]
  }'

import Anthropic from "@anthropic-ai/sdk"

const mizan = new Anthropic({
  baseURL: "https://api.app-mizan.com",
  apiKey: process.env.MIZAN_KEY, // "mizan_rt_..."
})

const message = await mizan.messages.create({
  model: "claude-sonnet-4-5",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Say hi in five words." }],
})

console.log(message.content)

from anthropic import Anthropic

mizan = Anthropic(
    base_url="https://api.app-mizan.com",
    api_key=os.environ["MIZAN_KEY"],  # "mizan_rt_..."
)

message = mizan.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Say hi in five words."}],
)

print(message.content)

Model routing

A bare model id (no provider prefix) — exactly what the Anthropic SDK sends by default — defaults to the anthropic provider, as in the examples above. Prefix the model with any other supported provider to route through the Anthropic wire format to a different one — the response still comes back Anthropic-shaped either way.

curl https://api.app-mizan.com/v1/messages \
  -H "x-api-key: mizan_rt_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.1",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Say hi in five words." }]
  }'

auto and mizan/switchboard work here too, exactly as documented for Auto Router and Switchboard.

Streaming

Set stream: true and Mizan streams real Anthropic SSE events — message_start, content_block_start/delta/stop, message_delta, message_stop— no matter which underlying provider serves the request. The Anthropic SDK's own stream helpers work unmodified.

curl https://api.app-mizan.com/v1/messages \
  -H "x-api-key: mizan_rt_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5",
    "max_tokens": 1024,
    "stream": true,
    "messages": [{ "role": "user", "content": "Count to five." }]
  }'

Tool use

tools and tool_choice are fully supported, both streaming and non-streaming — tool_use blocks in a response, and tool_result blocks on a follow-up turn. This works whether the request routes to Anthropic or to any other provider through this endpoint.

{
  "model": "claude-sonnet-4-5",
  "max_tokens": 1024,
  "tools": [
    {
      "name": "get_weather",
      "description": "Get the current weather for a city.",
      "input_schema": {
        "type": "object",
        "properties": { "city": { "type": "string" } },
        "required": ["city"]
      }
    }
  ],
  "messages": [{ "role": "user", "content": "What's the weather in Beirut?" }]
}

Cost & balance

Every response carries the resolved provider/model slug and per-call cost — same convention as /v1/chat/completions, just inside the standard Anthropic usage object instead of OpenAI's. The same X-Mizan-Cost-Usd/X-Mizan-Balance-After-Usd response headers are set too.

{
  "id": "msg_...",
  "type": "message",
  "role": "assistant",
  "model": "anthropic/claude-sonnet-4-5",
  "content": [{ "type": "text", "text": "Hi there, hope you're well!" }],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 12,
    "output_tokens": 9,
    "cost": 0.000267,
    "balance_after_usd": 47.499733
  }
}

Known limitations

Image content blocks (vision) and extended-thinking blocks aren't supported yet — a request containing either is rejected with a 400 rather than silently dropped, so you're never billed for content that wasn't actually forwarded. Your own cache_control hints are accepted but not preserved through the translation — Mizan applies its own automatic caching on top of your system prompt and tools regardless, see Prompt caching. Gemini's image-generation models are also rejected with a 400 here — they're only reachable through /v1/chat/completions.

Errors

Errors use Anthropic's own envelope shape — the same HTTP status codes documented for /v1/chat/completions, just wrapped differently so the Anthropic SDK's typed exceptions resolve correctly.

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "..."
  }
}