Responses

The OpenAI-style Responses endpoint (stateless), request parameters, and streaming events

POST /v1/responses is the OpenAI-style Responses endpoint. It serves clients that default to the Responses API, including the Vercel AI SDK and Codex-style agents, without code changes. It runs the same model and pipeline as Chat Completions.

This endpoint is stateless. Ablatic never stores conversation content server-side, so store: true, previous_response_id, and background: true are rejected with a 400. Send the full conversation in input on every request.

Request

Authenticate with Authorization: Bearer sk-ablatic-....

curl https://api.ablatic.ai/v1/responses \
  -H "Authorization: Bearer $ABLATIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "talos-preview",
    "input": "Write a haiku about Linz"
  }'

input is either a string or an array of items (messages, function_call, function_call_output).

Field	Type	Purpose
`model`	string	`talos-preview` (or `talos`)
`input`	string \| array	The prompt or the full item list
`instructions`	string	System instructions
`tools`	array	Function tools in the flat Responses shape `{type, name, description, parameters}`
`tool_choice`	string \| object	`auto` / `none` / `required` or `{type, name}`
`reasoning`	object	`{ "effort": "low" \| "medium" \| "high" }` selects the thinking depth
`text`	object	`{ "format": { "type": "json_schema", ... } }` for structured output
`max_output_tokens`	integer	Output cap, 1 to 200000
`temperature`, `top_p`	number	Sampling controls
`stream`	boolean	Stream Responses events over SSE

Response

A non-streaming call returns a response object. output holds a message item with an output_text part, plus one function_call item per tool call. usage reports input_tokens, output_tokens, and output_tokens_details.reasoning_tokens. store is always false.

Streaming

Set "stream": true. The endpoint emits the Responses event graph: response.created, response.in_progress, response.output_item.added, response.content_part.added, response.output_text.delta (repeated), the matching .done events, and response.completed carrying the final object and usage. There is no [DONE] sentinel. Tool calls stream as response.function_call_arguments.delta.

Reasoning

reasoning.effort selects the thinking depth. When the reasoning trace is enabled for your API key (toggle it in the console), the response returns it as a reasoning output item whose summary holds the trace text; streaming delivers it as response.reasoning_summary_text.delta events ahead of the answer. usage.output_tokens_details.reasoning_tokens reports the token count either way.

SDKs

openai (Python)

from openai import OpenAI

client = OpenAI(base_url="https://api.ablatic.ai/v1", api_key="sk-ablatic-...")
resp = client.responses.create(model="talos-preview", input="Hello")
print(resp.output_text)

Vercel AI SDK

import { createOpenAI } from "@ai-sdk/openai";
import { generateText } from "ai";

const ablatic = createOpenAI({ baseURL: "https://api.ablatic.ai/v1", apiKey: "sk-ablatic-..." });
const { text } = await generateText({ model: ablatic("talos-preview"), prompt: "Hello" });