Docs

Responses API

Updated 2026-08-29

/v1/responses is one of OpenAI’s primary native endpoints. OpenAI recommends evaluating Responses for new projects; if your client or framework defaults to Chat Completions, use OpenAI-compatible calls.

The Base URL remains https://api.rokoapi.com/v1.

Endpoint

POST /v1/responses

Quick Start

curl https://api.rokoapi.com/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "gpt-4o",
    "input": "Introduce yourself in one sentence",
    "instructions": "You are a concise assistant"
  }'

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_API_KEY",
    base_url="https://api.rokoapi.com/v1",
)

response = client.responses.create(
    model="gpt-4o",
    input="Introduce yourself in one sentence",
    instructions="You are a concise assistant",
)
print(response.output_text)

For text extraction, prefer the SDK’s output_text; when manually iterating over output , note that the first item may be reasoning rather than message.

Common Parameters

Parameter	Description
`model`	Model ID of an available model that supports Responses
`input`	String or array of messages
`instructions`	System instructions (similar to a system prompt)
`max_output_tokens`	Maximum output tokens
`stream`	Semantic event stream
`tools` / `tool_choice`	Functions and built-in tools (subject to upstream support)

Model capability labels

Model pages display the Responses capabilities configured in the admin console. Check these labels before sending a request:

Label	Meaning
`responses-native`	The upstream provider supports the Responses protocol natively
`responses-compatible`	The gateway converts the request to another upstream protocol; only declared capabilities are guaranteed
`structured-outputs`	JSON Schema structured outputs are supported
`previous-response-id`	`previous_response_id` can continue a response
`conversation-state`	Upstream-managed conversation state is supported
`responses-compact`	`POST /v1/responses/compact` is supported
`stored-responses`	The `store` option is supported
`parallel-tool-calls`	Parallel tool calls are supported
`max-tool-calls`	The `max_tool_calls` limit is supported

Clients must not assume capabilities that are not shown. Compatibility-converted models do not advertise features that depend on an upstream resource lifecycle, such as conversation state, stored responses, or compaction.

Current compatibility-converted models

Model	Conversion	Supported behavior
`deepseek-v4-flash`	Responses to and from DeepSeek Chat Completions	Text, function tools, tool results, and streaming events
`deepseek-v4-pro`	Responses to and from DeepSeek Chat Completions	Text, function tools, tool results, and streaming events
`glm-5.3`	Responses to and from Zhipu V4 Chat Completions	Text, function tools, tool results, and streaming events
`glm-5.3-flash`	Responses to and from Zhipu V4 Chat Completions	Text, function tools, tool results, and streaming events

These models use stateless compatibility conversion. Keep the full conversation and tool-call history on the client and send it again in the next input. They do not support previous_response_id, conversation, compact, or stored responses.

Multi-turn Conversations

By default, maintain conversation history on the client side and include the full context in the input array. Use stateful parameters only when the model page explicitly shows previous-response-id or conversation-state.

Streaming Output

Responses streaming uses semantic events (e.g., response.output_text.delta), unlike Chat Completions’ choices[0].delta . Set stream: true and handle events by type.

Comparison with Chat Completions

Chat Completions	Responses
`messages`	`input`
System message	`instructions`
`max_tokens`	`max_output_tokens`
`choices[0].message.content`	`output_text`

Notes

First, check the Models page or the available models documentation to confirm whether the target model supports Responses.
Built-in tools, background tasks, and similar capabilities depend on upstream support; capabilities that are not declared on the model page are not part of the compatibility contract.