Messages (Anthropic)

Token Factory is also reachable through Anthropic’s Messages API, so a client built for Claude, such as Claude Code or the anthropic SDK, works once you change the base URL and the key.

This surface belongs to the gateway in front of Token Factory. A model service on your instance is your own vLLM server, so whether /v1/messages works there depends on the vLLM version it runs.

POST /v1/messages API key (Bearer or x-api-key)

Also reachable at /api/v1/messages.

Model names come from this platform

Send an id from GET /v1/models, not an Anthropic model name. claude-3-5-sonnet-20241022 is not served here and will be rejected with 400.

Request

Parameter	Type		Description
model	string	Required	Model to run, from Token Factory.
messages	array	Required	Conversation so far, in Anthropic’s block format.
max_tokens	integer	Required	Cap on tokens generated. Anthropic requires this; so does the gateway.
system	string or array	Optional	System prompt, as a string or as an array of text blocks.
temperature	number	Optional	Sampling temperature, 0 to 1. This is Anthropic’s range, not OpenAI’s, so 1.5 is rejected with 400.
stream	boolean	Optional	Stream the response as server-sent events. Defaults to false.
tools	array	Optional	Tool definitions, if the model supports tool calling.
thinking	object	Optional	Extended-thinking configuration. Mapped onto the reasoning controls the backend understands. The budget_tokens form is currently rejected; see below.
output_config	object	Optional	effort controls adaptive reasoning depth on models that support it. Accepted tiers vary per model; low and medium work on every model whose tiers are verified.
metadata	object	Optional	user_id is used for sticky routing. Claude Code puts its session id here.

Unlisted parameters are dropped, not forwarded

As on chat completions, the request is validated against the set above and rebuilt before it reaches the backend. top_p, top_k and stop_sequences are not in that set, so they are removed silently.

Budget-based extended thinking is rejected

thinking: {"type": "enabled", "budget_tokens": N}, which Claude Code sends by default, is mapped onto reasoning.max_tokens. None of the models currently served support a token budget for thinking, so the request returns 400:

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "\"thinking\" is not supported for this model. Remove the \"thinking\" parameter or use a model that supports extended thinking."
  }
}

This is not an outage. The serving engine drops budget_tokens silently, so letting it through would only hand you an answer with no thinking in it.

To enable thinking on this endpoint, use output_config.effort:

{
  "model": "DeepSeek-V4-Flash",
  "output_config": { "effort": "high" },
  "messages": [{ "role": "user", "content": "..." }],
  "max_tokens": 1024
}

The thinking text arrives as thinking blocks in content, alongside the text ones.

effort becomes reasoning_effort internally, so the same per-model tier limits apply: the high above is fine on DeepSeek-V4-Flash but returns 400 on Qwen3.8-Flash-Next. Stick to low and medium for one code path across models; those two are accepted by every model whose tiers are verified. How MiniCPM-V-4.7-1B, MiniCPM-V-4.7-35B-A3B and Gemma-4-31B-IT treat effort has not been verified. The full table is on POST /v1/chat/completions.

Alternatively, use reasoning_effort on that endpoint.

Example

curl https://developer.amd.com.cn/radeon/api/v1/messages \
  -H "x-api-key: $RADEON_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "DeepSeek-V4-Flash",
    "max_tokens": 256,
    "system": "You are a concise assistant.",
    "messages": [
      {"role": "user", "content": "What is ROCm?"}
    ]
  }'

from anthropic import Anthropic

client = Anthropic(
    base_url="https://developer.amd.com.cn/radeon/api",
    api_key="rc-...",
)

message = client.messages.create(
    model="DeepSeek-V4-Flash",
    max_tokens=256,
    messages=[{"role": "user", "content": "Explain ROCm in two sentences."}],
)
print(message.content[0].text)

The SDK appends /v1/messages itself, which is why base_url stops at /api.

Response

{
  "id": "msg_9f3b21d0-4c8a-4f2e-b7d1-2a6c38e5b190",
  "type": "message",
  "role": "assistant",
  "model": "DeepSeek-V4-Flash",
  "content": [
    { "type": "text", "text": "ROCm is AMD's open software platform for GPU computing..." }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "usage": {
    "input_tokens": 24,
    "output_tokens": 118,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0
  }
}

stop_reason is one of end_turn, max_tokens, tool_use, or refusal. It is derived from the backend’s OpenAI-style finish_reason, and an unrecognised one becomes end_turn. stop_sequence is always null, because stop sequences aren’t forwarded in the first place. Reasoning models emit thinking blocks in content alongside the text ones.

Counting tokens

POST /v1/messages/count_tokens API key (Bearer or x-api-key)

Anthropic SDKs call this before sending, to size a request. Takes the same body as /v1/messages and returns:

{ "input_tokens": 24 }

system and tools are included in the count, because both are billed as input.

This is an estimate

The count comes from the gateway’s own tokenizer, not from the model that will serve the request. It will not match the usage.input_tokens you get back from /v1/messages, and it is not Anthropic’s counter either. Use it for sizing, not for billing. The authoritative numbers are in the usage block of the actual response.

Errors

Errors on both paths use Anthropic’s envelope, so SDK error handling works unchanged:

{
  "type": "error",
  "error": {
    "type": "authentication_error",
    "message": "Unauthorized: No API key provided."
  }
}

Refusals from the platform in front of the gateway, such as an invalid key or admission control, are wrapped in detail instead. See Errors.

Rate limits are shared with /v1/chat/completions: both land on the same backend capacity and count against the same per-key gates. See Rate limits.