Skip to content

LLM for Computational Social Science

Runnable guidance and examples for using LLM APIs in computational social science research, from a first API call to processing tens of thousands of text messages efficiently and at low cost.

Using LLMs for research can be as simple as asking the models questions through a chat box. Processing tens of thousands of text messages programmatically is harder: you must keep keys safe, get output you can parse, run queries in parallel, and control cost. This website walks through each of those steps.

What this site does not cover

This is a technical guide to interacting with LLMs from code. It does not explain how LLMs work, and it is not about AI agents. It also does not cover methodological questions such as validating model output against human labels or reporting results.

The core chapters use OpenAI's API. The later chapters show the same patterns with Anthropic, hosted open-source models, and models running on your own computer.

Most chapters are runnable Jupyter notebooks. Use the Open in Colab badge at the top of a page to try the code without installing anything. You can also clone the repository and run the notebooks locally. The source code can be found in the GitHub repository.

Start here

Read the API key page first, then go through the OpenAI chapters in order.

  • Basics

    Your first API call from Python: set up the client, send a prompt, read the response.

    Basics

  • Structured output

    Get answers back as validated Python objects instead of free text.

    Structured output

  • Async programming

    Send many requests at once so a large dataset takes minutes, not hours.

    Async programming

  • Batch processing

    Upload all your prompts in one file and pay half price for large jobs.

    Batch processing

Other providers

The same patterns work beyond OpenAI. Each chapter covers setup, a basic query, and structured output.

  • Anthropic API

    Claude models through Anthropic's Python package.

    Anthropic API

  • Open-source models

    Open-weight models through a hosted API (OpenRouter), using the same openai package.

    Open-source models

  • Local LLMs

    Run a model on your own computer with Ollama: no key, no per-token cost, data stays local.

    Local LLMs

Which provider should I use?

The tutorial covers four ways to reach a model. The code is nearly the same for all of them. The openai Python package works with every provider except Anthropic, whose package has a similar interface. Prices are per 1M tokens (input / output) for the example model each chapter uses, as of September 2026; check the provider's pricing page before a large run.

Provider Example model Price per 1M tokens Setup Pick it when
OpenAI gpt-5.6-luna $0.20 / $1.20 One account and key You are starting out and want the most capable ecosystem, the best documentation, or a batch API that halves the price for large jobs.
Anthropic claude-haiku-4-5 $1 / $5 One account and key You want a second model family to check robustness, Claude works better on your task, or you want a 50% batch discount.
Open-source models via OpenRouter deepseek/deepseek-v4-flash $0.08 / $0.15 One account and key You want the lowest per-token cost, access to many open-weight models through one key, or a model you can name exactly in a paper.
Local with Ollama gemma4:e2b Free Install Ollama, download a 7 GB model Your data cannot leave your machine, or you need to rerun the same model years later. Slower and less capable than the hosted options on a laptop; a GPU workstation runs 100B-class models.

A practical default: prototype the prompt on a few hundred examples with OpenAI, then decide. If the task is easy for the model, switch to the cheapest option that still passes your validation set. If the data is sensitive, start with the local chapter instead.

Dependencies

We use uv to manage dependencies for this project. To install them, clone the repository and run:

uv sync

Run a script with:

uv run script.py

Questions and contributions

If you have questions or suggestions, please open an issue. Pull requests are also welcome.

About

Created by Kaicheng Yang with help from Claude Code.

You may also find other tools from our lab useful: