OpenAI API

OpenAI API#

Chat vs. API#

OpenAI provides two separate services — make sure you’re using the right one:

  • ChatGPT (chat.openai.com) — web-based chat interface; not suitable for batch processing

  • OpenAI API (platform.openai.com) — programmatic access; pay-as-you-go; scalable to thousands of inputs

For research workflows, you want the API.

Setting Up#

1. Create an API Account#

Go to platform.openai.com and create an account (separate from your ChatGPT account).

2. Create an API Key#

In the API keys page :

  1. Click Create new secret key

  2. Copy and save it immediately — it will not be shown again

  3. Set a spending limit under Billing → Limits to avoid unexpected charges

3. Store Your Key Securely#

Warning

Never put your API key directly in source code or commit it to a GitHub repository.

The recommended approach is to store your key in a .env file and load it at runtime:

# ~/.env or /kellogg/proj/<your-netid>/keys/.env
OPENAI_API_KEY=sk-proj-...
OPENAI_ORG_ID=org-...

Load it in Python:

import os
from dotenv import load_dotenv  # pip install python-dotenv

load_dotenv()
api_key = os.getenv("OPENAI_API_KEY")

Alternatively, store the key in a plain text file and read it:

with open("/kellogg/proj/<your-netid>/keys/openai-key.txt", "r") as f:
    api_key = f.read().strip()

Making Your First API Call#

from openai import OpenAI  # pip install openai

client = OpenAI()  # reads OPENAI_API_KEY from environment

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "You are a helpful research assistant."},
        {"role": "user",   "content": "Summarize the key ideas in behavioral finance."}
    ],
    temperature=0,   # 0 = deterministic; increase for more variety
    max_tokens=512,
    seed=42          # improves reproducibility
)

print(response.choices[0].message.content)

Best Practices for Research#

These best practices align with the LLM API Usage on KLC recommendations.

Reproducibility

  • Set seed and document the exact model version (e.g., gpt-4o-2024-08-06)

  • Log all prompts, parameters, and responses — models can update and outputs can change

  • Save raw API responses before any post-processing

Cost Management

  • Set a max billing limit in your OpenAI account settings

  • Develop and test on small samples before running at scale

  • Use max_tokens to bound per-request cost

Data Privacy

  • Check your institution’s data governance policies before sending any data to OpenAI

  • For sensitive or IRB-governed data, use open-source models on KLC instead

Validation

  • LLMs can produce errors, hallucinations, and biases

  • Build unit tests that validate outputs on known examples before deploying at scale

  • See Validation and Rigor for a framework

The OpenAI Playground#

The Playground is a web-based interface for developing and testing prompts before writing code. Use it to:

  • Experiment with different system prompts

  • Compare models side by side

  • Fine-tune parameters (temperature, max tokens, top_p)

Monitoring Usage#

Track your token usage and costs at platform.openai.com/usage . Set alerts under Billing → Notification thresholds.