Preview, best effort: there is no uptime promise yet.
APIs
Chat completions (/v1/chat/completions), Anthropic’s Messages API (/v1/messages) and OpenAI’s Responses API (/v1/responses), streamed or not. Any other API naming these ids is refused with 400 catalog_api_unsupported.
The base URL is https://gateway.os3compute.com/v1 (Messages: https://gateway.os3compute.com). It answers from the nearer of OS3’s two US locations.
Context and output
Context: 1,048,512 tokens, prompt and answer together. A prompt that does not fit is refused, never cut.- By default an answer is at most 8,192 tokens without thinking, 65,536 with it and 131,072 at
maxeffort. Setmax_tokens(1 to 393,216) for another cap. Thinking tokens count toward it. - One completion a request (
n= 1).
Thinking
- The model thinks unless you send
reasoning_effort: "none"(orthinking: {"type": "disabled"}). - Efforts are
low,high(the default) andmax. - Thinking comes back in
reasoning_content, apart fromcontent. Its tokens are billed as output. - On a tool turn, send the assistant message back with its
reasoning_content. - Messages API: a
thinkingblock with an empty signature. Responses API: areasoningitem withreasoning_text.
Tools
- Your own function tools, with tool choice
auto,none,requiredor a named tool. - Tools a provider runs on its servers (web search, code execution, computer use and the like) are refused with
400 unsupported_tool.
Images
- Up to 600 a request, as data URLs or public
http(s)URLs. - Images inside a tool result work too.
Streaming and long answers
- A stream sends
: keep-alivelines while the model works (Messages:event: ping). - A long answer that is not streamed starts with status 200 after 10 seconds and sends a blank line every 10 seconds until its JSON is ready. The OpenAI and Anthropic SDKs read it as usual.
- Prompts above 200,000 tokens: rarely, one takes 10 to 30 seconds longer to start answering. Set a first-token timeout of at least 120 seconds.
Responses API
OS3 keeps no response: send the whole conversation each turn.store: true, previous_response_id, conversation and background are refused with 400 unsupported_parameter.
Answers
modelis the id you sent.system_fingerprintnames the serving configuration that answered.- Every answer carries an
x-os3-request-idheader. - Errors have OpenAI’s shape (Anthropic’s on the Messages API). See Errors.