Responses

POST https://api.inferam.ai/v1/responses is OpenAI’s Responses format, which Codex uses. Inferam translates it to the model server’s Chat Completions format and back, including streaming events.

curl https://api.inferam.ai/v1/responses \
  -H "Authorization: Bearer $INFERAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "DeepSeek-V4.1-Flash-1017",
    "instructions": "You are a terse senior engineer.",
    "input": "Explain Rust lifetimes in two sentences.",
    "reasoning": {"effort": "low"}
  }'

Supported fields

model, input (a string or a list of items), instructions, stream, max_output_tokens, temperature, top_p, tools, tool_choice, text.format (json_object, json_schema) and reasoning.effort.

• Input items: message (roles user, assistant, system, developer), reasoning, function_call, function_call_output, custom_tool_call and custom_tool_call_output.
• Content parts: input_text, output_text and input_image with a base64 data: URL.
• Tools: function tools, plus custom (freeform) tools such as Codex’s apply_patch. Hosted tools (web_search, file_search, computer_use, local_shell, …) can’t run on Inferam and are dropped from the request.

Stateless only

Inferam keeps no conversation state. store is always false, and a request with previous_response_id gets a 400. Send the whole conversation in input, as Codex does. Include the reasoning items from earlier turns: the model expects its own reasoning back on tool-calling turns.

Reasoning

reasoning.effort takes none (thinking off), minimal, low, medium, high, xhigh or max. The model’s reasoning comes back as a reasoning output item with the text in content (reasoning_text). When streaming, it arrives as response.reasoning_summary_text.delta events. usage.output_tokens_details.reasoning_tokens is always reported as 0; reasoning is counted in output_tokens.

Streaming

With "stream": true you get the standard event sequence: response.created, response.in_progress, the output item, content and function-call argument events, then response.completed. If the output limit is reached, the stream ends with response.incomplete instead. A failure after streaming has started ends with response.failed.

Usage

"usage": {
  "input_tokens": 2048,
  "input_tokens_details": { "cached_tokens": 1792 },
  "output_tokens": 312,
  "output_tokens_details": { "reasoning_tokens": 0 },
  "total_tokens": 2360
}