Create Anthropic Messages

OpenAI-compatible gateway endpoint for Anthropic Messages API payloads.

POST/v1/messages

Legacy path deprecated

The legacy aliases /v1/anthropic/messages and /v1/anthropic/messages/count_tokens are deprecated and will be removed on September 3, 2026. Responses on those paths carry Deprecation and Sunset headers. Migration is a base-URL change only: point your Anthropic SDK at https://api.therouter.ai — the SDK appends /v1/messages itself.

Request Parameters

Name	Type	Required	Description
model	string	Required	Anthropic model ID, for example anthropic/claude-sonnet-4.5.
messages	array	Required	Anthropic-formatted message list.
max_tokens	integer	Required	Maximum tokens to generate.
system	string \| array		System prompt string or array of system content blocks.
temperature	number		Sampling temperature between 0 and 1.
top_p	number		Nucleus sampling threshold.
top_k	integer		Top-K sampling: only consider the K most likely tokens.
stop_sequences	array		Sequences that will cause the model to stop generating.
stream	boolean		If true, returns a streaming SSE response. Defaults to false.
metadata	object		Arbitrary metadata object passed through with the request.
tools	array		Tool definitions available to the model (Anthropic tool format).
tool_choice	object		How the model selects tools: {"type": "auto"}, {"type": "any"}, or {"type": "tool", "name": "..."}.
thinking	object		Extended thinking controls, e.g. {"type": "enabled", "budget_tokens": 4096}. This endpoint takes the native Anthropic thinking block — the normalized reasoning object is an OpenAI-compatible-endpoint parameter (/v1/chat/completions) and is not translated here. On Claude 4.6+ models thinking is adaptive — the model self-selects its reasoning depth and budget_tokens does not apply.

Authentication

This endpoint accepts both Authorization: Bearer and x-api-key headers. Use x-api-key for drop-in compatibility with the native Anthropic SDK.

Request Example

bash

curl -X POST https://api.therouter.ai/v1/messages \
  -H "x-api-key: $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4.5",
    "max_tokens": 256,
    "messages": [{"role": "user", "content": "Explain retrieval augmented generation."}]
  }'

Response

json

{
  "id": "msg_abc123",
  "type": "message",
  "role": "assistant",
  "content": [{"type": "text", "text": "RAG combines..."}],
  "model": "anthropic/claude-sonnet-4.5",
  "stop_reason": "end_turn",
  "usage": {"input_tokens": 22, "output_tokens": 81}
}

Streaming

Set stream: true to receive the response as server-sent events (SSE). The gateway supports two routing paths, both producing Anthropic-format SSE events:

Name	Type	Required	Description
Native passthrough			When the resolved provider is Anthropic, events are streamed directly from the upstream API.
Translation path			When routing to a non-Anthropic provider, the gateway translates OpenAI SSE chunks into Anthropic-format events in real time.

Events follow the standard Anthropic streaming format:

text

event: message_start
event: content_block_start
event: content_block_delta     # repeated for each text chunk
event: content_block_stop
event: message_delta           # final usage and stop_reason
event: message_stop

The stream is opened as soon as the request has been routed, and until the first upstream event the gateway writes Anthropic's event: ping every 15 seconds (the first one immediately), so intermediaries do not close an idle connection while a reasoning model is thinking. The Anthropic SDKs already ignore ping. An error after the stream is open arrives as a single event: error.

Release note — pre-dispatch errors on stream: true (gateway release of 2026-10-11)

From the gateway release of 2026-10-11, a stream: true request to /v1/messages that fails before it is sent upstream — unknown model, a guardrails block, insufficient balance or quota — returns an HTTP error in the Anthropic error envelope, exactly as the Anthropic API does, instead of a 200 followed by event: error. SDK clients already handle this; a hand-written SSE consumer that assumed every streaming response is a 200 should check the status code first. In the same release, a stream that is not delivered in full is charged for the prompt plus the output actually written to your connection, with the reason recorded in the request log as delivery_outcome (values are listed in the Chat Completions reference); non-streaming requests remain subject to the edge's 100-second first-byte limit — use streaming for reasoning models.

Notes

Use this route when migrating existing Anthropic SDK payloads. Set your SDK base URL to https://api.therouter.ai (the SDK appends /v1/messages itself) and use your TheRouter API key.

Chat Completions

Analytics