Request Parameters, in Full

1 · The core three essential

Three parameters do the heavy lifting on every call, and you already know them. model picks which GPT (keep it in config — ox1). instructions is the system/developer role — the stable rules. input is the user content — a string for a one-shot, or a list of role/content turns for a conversation or multimodal request. Master these and you can make any call; the rest of this chapter is the dials you reach for on top of them.

core.pyresp = client.responses.create(
    model="gpt-5.5",                 # which model (from config)
    instructions="You are terse.",     # system/developer role
    input="Define idempotency.",       # user content (str or list)
)

2 · Reasoning effort & output length essential

Two dials control how hard the model thinks and how much it says. reasoning={"effort": …} (low/medium/high/max) sets how much hidden reasoning a reasoning-capable model does — more for hard problems, less for easy ones. max_output_tokens caps the reply length — a hard ceiling that both controls cost (output bills high) and prevents runaway generations.

effort.pyresp = client.responses.create(
    model="gpt-5.5",
    reasoning={"effort": "high"},     # low | medium | high | max
    max_output_tokens=800,            # hard cap on the reply
    input="Plan a zero-downtime database migration.",
)

3 · Sampling — temperature & top_p intermediate

These control randomness, and the honest advice is: usually leave them alone. temperature (roughly 0–2) scales how adventurous the next-token choice is — low is focused and near-deterministic, high is creative and varied. top_p (nucleus sampling, 0–1) instead limits the choice to the most-probable mass of tokens. The key discipline: tune one, not both, and only when you have a reason — low temperature for extraction/classification where you want stable, repeatable output; higher for brainstorming or creative writing where variety is the point.

The common mistake is reflexively setting temperature=0 "for determinism" and expecting byte-identical output every time. Low temperature makes output more consistent, not provably identical — and on reasoning models the effort dial often matters more than temperature anyway. Default to leaving sampling untouched unless a measured problem tells you to change it.

4 · Structured output & tools intermediate

Two parameters turn free text into something a program can trust. For structured output, use client.responses.parse(..., text_format=PydanticModel) and read resp.output_parsed (covered in ox2) — the response is constrained to your schema. For tools, pass tools=[…] — a mix of your function tools ({"type":"function",…}) and hosted tools ({"type":"web_search"} etc., oap5); tool_choice can force or forbid tool use when you need to.

structured_and_tools.py# structured output
r = client.responses.parse(model="gpt-5.5", input="...", text_format=MyModel)
obj = r.output_parsed

# tools (function + hosted), optionally forced
r = client.responses.create(
    model="gpt-5.5", input="...",
    tools=[MY_FUNCTION_TOOL, {"type": "web_search"}],
    tool_choice="auto",              # auto | required | none
)

5 · State & metadata advanced

A few parameters manage conversation state and bookkeeping. store (bool) controls whether OpenAI retains the response server-side so it can be referenced later; previous_response_id can chain a new call onto a stored prior response, letting the platform carry context instead of you re-sending the whole transcript. metadata attaches your own key/value tags to a response — handy for correlating calls with a trace_id or feature label in your logs (oap4). These are optional conveniences; a basic stateless app ignores them and just re-sends its input list each turn.

6 · Quick reference professional

7 · Professional — a sane default profile professional

The professional move is to not re-decide every parameter on every call. Wrap responses.create in a small helper that encodes your defaults — model from config, a sensible max_output_tokens cap, effort chosen per task-type, sampling left untouched, and metadata auto-stamped with a trace_id and feature label — and let call sites override only what's genuinely different. That turns "which parameters?" from a per-call question into a one-time policy, and makes a parameter change (a model bump, a new cap) a one-line edit instead of a find-and-replace.

Knowledge check check yourself