Instagram

Docs

Responses API

The OpenAI Responses API, with stateful multi-turn conversations and Alvin auto-routing.

The Responses API is OpenAI's newer inference interface. It works across every provider in the catalog — pass any model ID, or elyxir-alvin to let Alvin route.

Endpoint

POST https://api.elyxir.ai/v1/responses

How it differs from Chat Completions

	Chat Completions	Responses
Prompt field	`messages` array	`input` — string or array
System prompt	A `system` message	`instructions`
Multi-turn	Resend the full history	`previous_response_id`
Output cap	`max_tokens`	`max_output_tokens`
Streaming	Token deltas	Richer typed events

The big practical win is previous_response_id: you continue a conversation by referencing the last response instead of resending every prior turn.

Body parameters

Parameter	Type	Required	Description
`model`	string	Yes	Model ID, or `elyxir-alvin` for auto-routing
`input`	string \| array	Yes	A prompt string, or an array of `{role, content}` items
`instructions`	string	No	System-level instructions
`temperature`	number	No	Defaults to 0.7
`max_output_tokens`	integer	No	Cap on generated tokens
`stream`	boolean	No	Stream the response. Defaults to false
`previous_response_id`	string	No	Continue from an earlier response
`tools`	array	No	Custom function tools — see below

Roles accepted in a structured input are developer, user, and assistant.

Examples

curl -X POST https://api.elyxir.ai/v1/responses \
  -H "Authorization: Bearer elyxir_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "elyxir-alvin",
    "input": "Write a Python function to sort a list using merge sort",
    "max_output_tokens": 2000
  }'

With system instructions

{
  "model": "claude-sonnet-4-6",
  "instructions": "You are a senior software engineer. Be concise and include code examples.",
  "input": [
    {"role": "user", "content": "How do I implement retry logic with exponential backoff?"}
  ],
  "temperature": 0.3
}

Continuing a conversation

Reference the previous response instead of resending the history:

{
  "model": "gpt-4o",
  "input": "Can you add error handling to that function?",
  "previous_response_id": "resp_abc123"
}

Tools

Custom function tools work as they do in the OpenAI API — send type: "function" entries in tools.

The built-in tool types (web_search, file_search, code_interpreter) are not supported here. If you send them they are stripped before the request reaches the model, silently — the call succeeds without them. For web-grounded answers use a Perplexity Sonar model, or enable web search in Alvin Studio.

Alvin routing

Set model to elyxir-alvin and the response carries the routing decision in X-Alvin-Model, X-Alvin-Task-Type and X-Alvin-Confidence headers, plus a routing field in the body. See Alvin routing.

Which should I use?

Use Chat Completions if you have existing OpenAI-compatible code, or you want the widest client-library support.

Use Responses for new work, especially long multi-turn conversations where previous_response_id saves you resending — and paying for — the whole history on every turn.

Chat Completions — the widely-supported alternative
Models — discover available model IDs
Error handling — status codes and retries

Chat Completions Models