OpenAI SDK: Setup, keys & your first call

Setup takes five minutes: a virtual environment, the openai package, and an API key in an environment variable. The one rule that matters: an API key is a password with a credit card attached. It never goes in code, notebooks, Git or chat messages.

1. Install

python -m venv .venv
source .venv/bin/activate            # Windows PowerShell: .venv\Scripts\Activate.ps1
pip install openai python-dotenv     # (uv users: uv add openai python-dotenv)
python -c "import openai; print(openai.__version__)"

This guide was written against openai 2.44 (September 2026). Minor versions add features often; check the SDK changelog when something here does not match.

2. The API key

Environment variable

export OPENAI_API_KEY="sk-..."                  # macOS/Linux, current shell only
setx OPENAI_API_KEY "sk-..."                    # Windows, persists for new terminals

.env file

.env (add .env to .gitignore FIRST)

OPENAI_API_KEY=sk-...

from dotenv import load_dotenv
load_dotenv()                  # copies .env into os.environ; OpenAI() then finds the key

3. Your first call

hello.py (real API: needs OPENAI_API_KEY)

from openai import OpenAI

client = OpenAI()      # reads OPENAI_API_KEY (and optional OPENAI_PROJECT, OPENAI_BASE_URL) from the environment
response = client.responses.create(
    model="gpt-5.6-luna",                       # check the current models page; names change
    input="Explain what a data lakehouse is in two sentences.",
)
print(response.output_text)
print(response.usage)

4. Running every example without a key: fake_openai.py

Download openai-sdk/fake_openai.py from this site’s repository and put it next to your script. It creates a normal OpenAI client whose HTTP traffic goes to a small local fake server, so the real SDK code runs (request building, retries, parsing, streaming, error types) with scripted replies. That is how every example in this guide was executed, and why the outputs you see are real SDK output.

from fake_openai import fake_client

client = fake_client(script=["A lakehouse stores data as open files in cheap object storage "
                             "and adds transactions and governance on top."])
response = client.responses.create(model="gpt-5.6-luna",
                                   input="Explain what a data lakehouse is in two sentences.")
print(response.output_text)
print(response.usage.input_tokens, response.usage.output_tokens)   # the fake estimates ~4 chars/token
print(type(client).__name__, client.base_url)

Output

A lakehouse stores data as open files in cheap object storage and adds transactions and governance on top.
12 26
OpenAI https://fake.openai.local/v1/

5. Client options worth setting

from fake_openai import fake_client

client = fake_client(timeout=20.0, max_retries=3)   # the same keyword arguments OpenAI(...) accepts
print(client.timeout, client.max_retries)

strict = client.with_options(timeout=5.0, max_retries=0)   # per-call overrides, no new client
print(strict.timeout, strict.max_retries)

Output

20.0 3
5.0 0

Create one client per application and reuse it, because it holds a connection pool. The defaults are a 10-minute timeout and 2 retries. Most applications want a shorter timeout (part 12).

Recap