Messages API (Anthropic format)

POST https://paraloncloud.com/v1/messages accepts the request Claude Code and the Anthropic SDKs send and answers in the Message shape, served by the same open models as chat completions. A client written for Claude works by changing the base URL and the key, with no router in between.

Claude Code

Set these and start claude:

export ANTHROPIC_BASE_URL=https://paraloncloud.com
export ANTHROPIC_AUTH_TOKEN=sk-your-key
export ANTHROPIC_MODEL=qwen3.8-27b
export ANTHROPIC_SMALL_FAST_MODEL=qwen3.8-27b
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3.8-27b
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3.8-27b
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3.8-27b
claude

All the model variables matter. Claude Code sends side requests, such as a title for the session, to models other than the main one. With only the first three set we saw one of them go out under a Claude model name and come back "model not found"; with all five, every request of the session asked for qwen3.8-27b and was answered.

What we measured with Claude Code 2.1.296 on 13 October 2026: a folder with a slugify.js that has a bug and three tests, two failing, and a prompt asking it to run the tests, find the bug, fix it and run them again. Three runs, all fixed, 3 of 3 tests passing, in 5 to 8 turns and 27 to 55 seconds. That is one short task; a long session has not been measured. The runs, what Claude Code sends and the two requests we refused on the first day are written up in Claude Code on an open model, no router.

Two things to know:

• The context window is 131,072 tokens. Claude Code does not know this model and assumes more. In a long session it may be refused before it thinks to compact; compact earlier by hand with /compact.
• Ignore the cost Claude Code prints. It has no price list for the model. What was charged is on Inference usage.

Anthropic SDK

import anthropic

client = anthropic.Anthropic(base_url="https://paraloncloud.com", api_key="sk-your-key")
message = client.messages.create(
    model="qwen3.8-27b",
    max_tokens=200,
    system="You answer in one sentence.",
    messages=[{"role": "user", "content": "What is a KV cache?"}],
)
print(message.content[0].text)

The key is read from x-api-key, the header the SDK sends, or from Authorization: Bearer.

What is supported

system	a string, or text blocks
Content blocks	text, image (base64 or URL, on vision models), tool_use, tool_result
Tools	tools with a name and an input_schema, which your client runs; tool_choice auto, any, tool, none
Sampling	max_tokens, temperature, top_p, top_k, stop_sequences
Streaming	message_start, content_block_start / content_block_delta / content_block_stop, message_delta, message_stop
Thinking	off unless the request asks for it (thinking.type enabled or adaptive); the reasoning then comes back in thinking blocks
Usage	input_tokens, output_tokens, and cache_read_input_tokens for what the worker read from its prompt cache

stop_reason is end_turn, max_tokens or tool_use.

cache_control is accepted and ignored: prompt prefixes are cached automatically, and a repeated prefix shows up in cache_read_input_tokens without being marked.

What is not

Each of these is answered with a 400 that names the field:

• Server tools (web search, code execution, computer use): they run on Anthropic's side. Tools your client runs itself work.
• document blocks. Send the text.
• POST /v1/messages/count_tokens does not exist.

Errors come in Anthropic's envelope: {"type": "error", "error": {"type": "...", "message": "..."}}.

Requests are metered and billed as chat requests on the same model, at the same price.