Anthropic Messages

Create Anthropic-compatible messages with native response and streaming event shapes.

`POST /v1/messages`

Creates a message using the Anthropic protocol. The incoming protocol controls the request, response, stream, and error shapes; it does not restrict routing to Anthropic-owned models.

POST /v1/messages
x-api-key: $API_KEY
anthropic-version: 2023-06-01
Content-Type: application/json

Authorization: Bearer $API_KEY is also accepted. The platform controls the protocol version used for the selected upstream endpoint.

Request body

Field	Type	Required	Description
`model`	string	Yes	Complete public model ID.
`messages`	array	Yes	Non-empty sequence of `user` and `assistant` messages.
`max_tokens`	integer	Yes	Maximum tokens to generate.
`system`	string or array	No	System prompt or system content blocks.
`stream`	boolean	No	Return Anthropic SSE events when `true`.
`temperature`	number	No	Sampling temperature supported by the model.
`top_p`	number	No	Nucleus sampling value.
`top_k`	integer	No	Model-specific top-k sampling value.
`stop_sequences`	string[]	No	Custom stop sequences.
`tools`	array	No	Anthropic function tool declarations.
`tool_choice`	object	No	`auto`, `any`, `none`, or a named tool choice.
`metadata`	object	No	Provider-compatible request metadata.

Message content can be a string or an array of content blocks. Common block types are text, image, tool_use, and tool_result. URL and base64 image sources are accepted when the selected model supports vision.

Sampling on current Claude models

Claude 4.7 and later models, and Claude Mythos Preview, reject non-default temperature, top_p, and top_k. Omit all three for those models and use prompting to control behavior.

Tool choice and prefill on current Claude models

Claude Opus 5.5, Claude Sonnet 5.5, Claude Fable 5.1, and Claude Mythos 5.1 reject forced tool use (tool_choice of any or a named tool) with a 400 error; use auto instead. Thinking is on by default on Claude 5 models, including Claude Haiku 5.5, and a prefilled assistant turn is rejected while thinking is on, so end messages with a user turn.

Request example

curl "http://127.0.0.1:11113/v1/messages" -H "x-api-key: $API_KEY" -H "anthropic-version: 2023-06-01" -H "Content-Type: application/json" -d '{
  "model": "anthropic/claude-sonnet-5-5",
  "max_tokens": 1024,
  "system": "Answer concisely.",
  "messages": [
    {"role": "user", "content": "What is a model gateway?"}
  ]
}'

Synchronous response

{
  "id": "msg_01J...",
  "type": "message",
  "role": "assistant",
  "model": "anthropic/claude-sonnet-5-5",
  "content": [
    {
      "type": "text",
      "text": "A model gateway provides one API over multiple AI models."
    }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "stop_details": null,
  "usage": {
    "input_tokens": 12,
    "output_tokens": 13
  }
}

The content field is always an array. Tool requests use tool_use blocks. Reasoning-capable responses can include thinking blocks when the selected model and route expose them. Native refusal responses use stop_reason: "refusal" and can include stop_details with the policy category.

Streaming response

Set stream: true to receive Anthropic event names and payloads. A normal stream follows this lifecycle:

message_start
One or more content_block_start, content_block_delta, and content_block_stop groups
message_delta with the stop reason and final usage
message_stop

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

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

Compatibility and restrictions

Same-protocol routes preserve Anthropic fields such as thinking, cache controls, metadata, and provider extensions. Cross-protocol conversion maps system instructions, text and image blocks, tool use and results, function tools, tool choice, output limits, temperature, top_p, stop sequences, stream events, and usage. Nonportable fields can be omitted and reported in x-dropped-params.

Anthropic content blocks have no compatible field for Gemini thought signatures. Multi-turn Gemini function calling through the Messages dialect can therefore fail when the upstream requires that signature; use Gemini Generate Content or another native-compatible route for that workflow.

Missing model, empty messages, or missing max_tokens returns 400 invalid_request. /v1/messages/count_tokens is not exposed. Client routing fields, paid service-tier selections, and inference_geo values other than global are rejected; global is accepted and removed before forwarding.