Testing OpenAI integrations without paying per call

Testing code that calls the OpenAI API gets expensive fast. Every iteration costs money, you burn through rate limits, and CI pipelines need a real API key to run. So I built a small library that fakes the OpenAI API: openai-api-mock.

It works with anything that speaks the OpenAI API since the mock intercepts by base URL, it also covers OpenAI-compatible providers like Azure OpenAI, Groq, Together, or a local Ollama instance.

The problem it solves

While building a project that used function calling, I had to test structured responses over and over. Each run cost money, and I couldn't run tests in CI without exposing a key. I wanted something I could drop into a test setup and forget about, so I wrote it.

Usage

nstall it as a dev dependency:

npm install -D openai-api-mock

Then call one function before your code makes OpenAI requests:

const { mockOpenAIResponse } = require('openai-api-mock');
mockOpenAIResponse(); // from here on, every OpenAI API call returns a mock response

Under the hood it uses nock to intercept HTTP requests to the OpenAI API and @faker-js/faker to generate the fake data.

What's new in 0.4.1

Version 0.4.1 adds support for the embeddings endpoint — including deterministic vectors (same input → same output), all input types, and both floatand base64encoding formats. This means you can test semantic search, RAG pipelines, and clustering with reproducible results.

// Request base64-encoded embeddings
const response = await openai.embeddings.create({
  model: 'text-embedding-3-small',
  input: 'test',
  encoding_format: 'base64',
});
console.log(response.data[0].embedding); // base64 string

If you hit a case it doesn't handle, open an issue on the repo. Feedback welcome.

Github Repo