Openai API

OpenAI API

Badass reference for wiring up the OpenAI API — built from the latest OpenAI docs (Responses API generation, migrated from Chat Completions and its being deprecated on a timeline; images/vision; audio; structured outputs; tools; pricing).

Golden rule: use the Responses API for anything new

Responses API (/v1/responses) is the recommended primitive for all new projects. Chat Completions (/v1/chat/completions) still works and is still supported, but Responses is the future-facing API: better performance with reasoning models, native multi-tool agentic loop in a single request, better prompt caching (lower cost), stateful conversation management, and it's what all new built-in tools (web search, file search, remote MCP, computer use, image generation, tool search) are designed around.

Only reach for Chat Completions when:

• You're maintaining/extending an existing Chat Completions integration and a full migration isn't worth it right now
• You need the audio-in-chat pattern (modalities: ["text","audio"]) — Responses docs currently describe text/image in, text out only; use Chat Completions with an audio-capable model (gpt-audio-1.5) for that specific case, or Realtime for live voice.

If the user has existing Chat Completions code and wants to modernize it, see references/migrating-to-responses.md.

Routing table — which reference file to read

Read only what's relevant — don't load every reference file for a simple question.

Quick-start pattern (Responses API, TypeScript)

import OpenAI from "openai";
const client = new OpenAI(); // reads OPENAI_API_KEY from env

const response = await client.responses.create({
  model: "gpt-5.6",           // pick current flagship unless user specifies otherwise
  instructions: "You are a helpful assistant.", // system-level guidance, highest priority
  input: "Hello!",            // string, or array of {role, content} items
});

console.log(response.output_text); // convenience aggregator of all text output

from openai import OpenAI
client = OpenAI()

response = client.responses.create(
    model="gpt-5.6",
    instructions="You are a helpful assistant.",
    input="Hello!",
)
print(response.output_text)

Gotcha: response.output is an array of typed Items (message, reasoning, function_call, function_call_output, etc.) — never assume the text lives at output[0].content[0].text. Use output_text for the common case, or iterate output by .type when you need reasoning/tool items too.

Gotchas & non-obvious constraints

• instructions doesn't carry across previous_response_id chains. Resend it every request if you're chaining responses instead of manually replaying input.
• Billing on chained responses: even with previous_response_id, all prior input tokens in the chain are billed as input tokens again — it's not free just because OpenAI is holding state for you.
• Structured Outputs schema location moved: response_format (Chat Completions) → text.format (Responses). Don't mix them up when porting code.
• Function/tool definitions are tagged differently: externally tagged in Chat Completions, internally tagged in Responses. Also, Responses attempts strict: true by default and silently falls back to non-strict if the schema can't be made strict-compatible (returns strict: false on the resolved tool) — check for this if you expected strict mode.
• Reasoning models (gpt-5.6, o-series) perform meaningfully better on Responses than Chat Completions — if the user is using a reasoning model, steer them to Responses even if they didn't ask about migrating.
• Starting with GPT-5.4, tool calling is not supported in Chat Completions with reasoning: none.
• Stateful by default: Responses (and Chat Completions for new accounts) store data by default. Set store: false to disable. For Zero Data Retention orgs, pair store: false with include: ["reasoning.encrypted_content"] to keep reasoning continuity without persistence — OpenAI enforces store: false automatically for ZDR orgs.
• Assistants API is deprecated (sunset Aug 26, 2026) — steer any Assistants API question toward Responses instead.
• n (multiple parallel generations) doesn't exist in Responses — one generation per request; fire multiple requests if you need candidates.
• Only gpt-5.4 and later support the tool_search deferred-tool-loading tool.
• Model names in these docs (e.g. gpt-5.6, gpt-5.6-sol/terra/luna, o4-mini) reflect the source docs' snapshot — always sanity-check current model availability against live docs if the user's request is cost- or capability-sensitive, since OpenAI ships new snapshots frequently.