Responses API

POST /api/v1/responses forwards native Responses requests through Portkey to the configured inference backend. Saved Portkey routing may select internally hosted models or OpenRouter/provider fallback. Model availability and optional capabilities are enabled only after model-specific compatibility and privacy tests pass. Requests remain native Responses throughout; there is no Responses-to-chat conversion. No server-managed response history is retained. Check supports_responses on /api/v1/models for enabled models.

Call /api/v1/models with your API key (Authorization: Bearer YOUR_IO_API_KEY) and select a model where supports_responses is true. Availability is account-specific during the restricted rollout. If no models have this flag enabled, your account does not currently have Responses access. Replace ENABLED_MODEL below with an enabled model ID from that list.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_IO_API_KEY",
    base_url="https://api.intelligence.io.solutions/api/v1",
)
response = client.responses.create(
    model="ENABLED_MODEL", input="Explain token caching.", store=False,
)
print(response.output_text)

Set stream=True for Responses SSE events. Events carry increasing sequence_number values, stable item IDs and a native terminal completed/incomplete/failed response. There is no [DONE] sentinel.

Supported contract

Feature	Behavior
Input	String or text message items; system/developer/user/assistant roles; instructions
Functions	Flat function definitions, auto/none/required/named choice, parallel calls
Structured output	`text.format`: text, json_object or json_schema when supported
Reasoning	Native effort controls when enabled; native reasoning-item replay; engine-supported encrypted content
Usage	Input/output/total tokens; cached/reasoning tokens; cache writes only when measured
Advisory fields	Metadata, client_metadata, user, safety_identifier, prompt_cache_key, service_tier, reasoning.summary and reasoning.context accepted but dropped; metadata returned empty, tier default
Stateless controls	`store:false`, no conversation or previous_response_id; truncation disabled

Tool loop / multiple turns

Clients resend previous input and output items, including exact function argument strings and call IDs. Tools execute in your application.

tools = [{"type": "function", "name": "weather",
          "parameters": {"type": "object", "properties": {
              "city": {"type": "string"}}, "required": ["city"]}}]
history = [{"role": "user", "content": "Weather in Paris?"}]
first = client.responses.create(model="ENABLED_MODEL", input=history,
                                tools=tools, tool_choice="required", store=False)
call = next(item for item in first.output if item.type == "function_call")
history += [item.model_dump(exclude_none=True) for item in first.output]
history += [{"type": "function_call_output", "call_id": call.call_id,
             "output": "Sunny, 20 C"}]
second = client.responses.create(model="ENABLED_MODEL", input=history,
                                 tools=tools, store=False)

Resolve every function call before replaying the next turn. previous_response_id is unsupported.

Compatibility limits

Unsupported parameters/tools return HTTP 400 with an OpenAI-style error object. No retrieve/delete/cancel/input-items routes. Storage, background execution, automatic truncation, image/file/audio/video input, hosted/custom tools, stored prompts, routing extensions, reasoning enabled/budget controls, logprobs and text verbosity are unsupported. include:["reasoning.encrypted_content"] and native encrypted replay are forwarded; availability depends on the backend. No encryption is synthesized. Accounts requiring storage receive HTTP 403. Unknown retention policy receives HTTP 503. Requests are limited to 2 MiB input, 128 tools/output items and 4 MiB upstream output. Explicit max_output_tokens values are capped at the model output limit; omitted values stay omitted. Transport has a 1800s total budget and streaming a 120s idle limit. After disconnect, the server keeps consuming within these bounds to collect measured usage. Missing or malformed required usage is never billed as zero. Billing persistence failures do not suppress completed output. OpenAI responses.create/responses.stream and the OpenRouter benchmark harness are compatibility targets. Some clients require unsupported parameters, tools or state. Codex 0.160.1 function-tool mode is tested with the configuration below; hosted/custom/namespace tools remain unsupported. Engine versions differ in event and nullable-field behavior. A model that fails the SDK or harness checks remains disabled. Missing or malformed final usage cannot be treated as zero; an unexpected stream end emits an error rather than an invented completion. Quota, context-limit and rate-limit errors retain actionable codes with sanitized messages; upstream authentication/routing failures return 502. Provider costs are not exposed in response usage. Stateless request handling does not certify downstream telemetry/storage retention. See the endpoint reference.

Codex 0.160.1

Codex 0.160.1 function-tool mode works with the configuration below. Use a model enabled for Responses and function tools.

model = "XiaomiMiMo/MiMo-V2.6-Flash"
model_provider = "io"
model_reasoning_effort = "none"
web_search = "disabled"

[agents]
enabled = false

[apps._default]
enabled = false

[model_providers.io]
name = "IO Intelligence"
base_url = "https://api.intelligence.io.solutions/api/v1"
env_key = "IO_API_KEY"
wire_api = "responses"

Set IO_API_KEY in your environment. App/agent namespaces and hosted search are disabled because this endpoint accepts function tools only. Other model presets and MCP plugins can introduce unsupported tools and need separate testing. Reasoning summary/context hints are accepted and ignored; Responses Lite behavior is not implemented. Omit max_output_tokens to let the engine select its budget; explicit values are capped at the model output limit. metadata retains its 16-pair/64-character-key/512-character-value limits; client_metadata is discarded without these limits. The overall request limit is 2 MiB. On models without reasoning enabled, reasoning.effort="none" is accepted and omitted upstream; active effort values remain unsupported. Unsupported tool errors include the tool type, Codex configuration keys and this guide link.

Getting Started with IO Intelligence

IO Intelligence Payments

Confidential Inference

Integrations

Supported contract

Tool loop / multiple turns

Compatibility limits

Codex 0.160.1

​Supported contract

​Tool loop / multiple turns

​Compatibility limits

​Codex 0.160.1

Supported contract

Tool loop / multiple turns

Compatibility limits

Codex 0.160.1