Responses

POST /v1/responses — OpenAI Responses API compatible endpoint for Codex CLI and other Responses API clients

Create a model response using the OpenAI Responses API shape. Drop-in replacement for the OpenAI Responses API — designed for Codex CLI and any client that speaks the Responses format.

Endpoint:

POST /v1/responses

Auth: Bearer token or x-api-key. Requires scope inference.chat.

Melious is stateless. store is always false, previous_response_id is rejected, and GET/DELETE /v1/responses/{id} return 404. Include the full conversation history in every request.

Example

from openai import OpenAI

client = OpenAI(
    api_key="sk-mel-<YOUR_API_KEY>",
    base_url="https://api.melious.ai/v1",
)

response = client.responses.create(
    model="glm-5.2",
    input="Name three Hanseatic cities.",
)
print(response.output_text)

curl https://api.melious.ai/v1/responses \
  -H "Authorization: Bearer sk-mel-<KEY>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.2",
    "input": "Name three Hanseatic cities."
  }'

export OPENAI_API_KEY="sk-mel-<YOUR_API_KEY>"
export OPENAI_BASE_URL="https://api.melious.ai/v1"
codex  # Works transparently

Request

Core parameters

Parameter	Type	Default	Description
`model`	string	—	Model ID, optionally with a flavor suffix like `:eco` or `:free`. See Routing.
`input`	string \| array	—	The conversation input. String for a single user message, or a list of input items (see below).
`instructions`	string	none	System-level instructions. Prepended as a system message. Equivalent to the system role in Chat Completions.
`max_output_tokens`	integer	model max	Caps the completion length. Maps to `max_tokens` internally.
`temperature`	number	model default	Sampling temperature, `[0, 2]`.
`top_p`	number	`1`	Nucleus sampling cutoff, `[0, 1]`.
`frequency_penalty`	number	`0`	Penalize repeated tokens, `[-2, 2]`.
`presence_penalty`	number	`0`	Penalize tokens already present, `[-2, 2]`.
`stop`	string \| array	null	Stop sequences.
`seed`	integer	none	Deterministic sampling (best-effort — not all providers honor it).
`user`	string	none	End-user identifier for abuse monitoring.

Stateless gateway

Parameter	Type	Default	Description
`store`	boolean	`false`	Accepted but always treated as `false`. Melious does not store responses server-side.
`previous_response_id`	string	—	Rejected with 400. Include the full conversation history in `input` instead.

Streaming

Parameter	Type	Default	Description
`stream`	boolean	`false`	Enable SSE streaming with Responses-format events.

See Streaming for the general streaming shape. Events are converted to native Responses API SSE format (response.output_text.delta, response.completed, etc.).

Reasoning

Parameter	Type	Default	Description
`reasoning`	object	none	`{ "effort": "low" \| "medium" \| "high" }`. Controls reasoning depth on reasoning models. Ignored by non-reasoning models.

Tools

Parameter	Type	Default	Description
`tools`	array	null	Tool definitions the model may call. Supports `function`, `file_search`, `web_search`, `web_search_preview`, `shell`, `function_shell`, `apply_patch`.
`tool_choice`	string \| object	`"auto"`	`\"auto\"`, `\"none\"`, `\"required\"`, or a specific tool reference.
`parallel_tool_calls`	boolean	`true`	Whether the model may call multiple tools in one response.

Tool types computer_use, code_interpreter, and mcp are not supported — they require server-side execution environments Melious does not host. Requests containing them are rejected with a validation error.

`input` items

When input is a list, each item is one of:

Message item:

{
  "type": "message",
  "role": "user",
  "content": "What's in this image?"
}

content can be a string or an array of content parts (text, image). Image parts use the same image_url format as Chat Completions — public URLs and base64 data URIs are both accepted.

Function call item (assistant tool call):

{
  "type": "function_call",
  "call_id": "call_abc",
  "name": "get_weather",
  "arguments": "{\"city\": \"Hamburg\"}"
}

Function call output item (tool result):

{
  "type": "function_call_output",
  "call_id": "call_abc",
  "output": "{\"temp\": 22}"
}

Item reference: { "type": "item_reference", "id": "..." } — rejected (no server-side state).

Response

{
  "id": "resp-...",
  "object": "response",
  "created_at": 1699999999,
  "model": "glm-5.2",
  "status": "completed",
  "output": [
    {
      "type": "message",
      "id": "msg_...",
      "role": "assistant",
      "content": [
        { "type": "output_text", "text": "Hamburg, Lübeck, Bremen." }
      ],
      "status": "completed"
    }
  ],
  "usage": {
    "input_tokens": 12,
    "output_tokens": 7,
    "total_tokens": 19
  },
  "store": false,
  "environment_impact": {
    "energy_kwh": 0.00015,
    "carbon_g_co2": 0.06,
    "water_liters": 0.0002,
    "renewable_percent": 85,
    "pue": 1.18,
    "provider_id": "ovhcloud",
    "location": "FR"
  },
  "billing_cost": {
    "energy": "0.0008",
    "credits": "0.0",
    "paid_with": "energy"
  }
}

`output`

Array of output items. The most common is a message item with content containing output_text parts. For tool calls, function_call items appear in the array.

`status`

completed — model finished normally.
failed — provider error or content filter.

`environment_impact` and `billing_cost`

Same Melious-specific fields as Chat completions. See Environmental impact and Pricing.

Stateless behavior

Feature	OpenAI	Melious
`store: true`	Stores response server-side	Ignored — always `false`
`previous_response_id`	Resumes from stored conversation	Rejected with 400
`GET /v1/responses/{id}`	Retrieves stored response	Returns 404
`DELETE /v1/responses/{id}`	Deletes stored response	Returns 404

Send the full conversation as input on every call. This matches how Codex CLI and most Responses API clients operate in practice.

Errors

Every error returns the standard {"error": {"code", "message", "details"}} shape. Common codes on this endpoint:

VALIDATION_4002 — model or input missing.
VALIDATION_4001 — previous_response_id provided (stateless gateway rejects it).
INFERENCE_3001 — unknown model ID.
INFERENCE_3207 — input exceeds the model's context window.
INFERENCE_3103 — all providers failed (transient; retry).
BILLING_2001 / BILLING_2003 — out of energy / credits.
AUTH_1015 — key is missing inference.chat scope.

Full list and retry guidance: Errors.

Chat completions • Streaming • Tool calling • Routing • Codex CLI integration

Responses

On this page