Groq, Not Grok: The Free AI API I Build Everything On

First, the spelling, because it trips everyone up. Grok with a K is the chatbot from xAI. Groq with a Q is a company that runs open models very, very fast and gives developers an API to use them. This article is about the second one, and about why almost everything I build now starts on its free tier.

Why I switched#

For a long time I used Gemini API keys for everything. It worked, until it didn't: the free limits kept running out in the middle of building, and then you just sit there and wait. Nothing kills a good evening of building like a quota error.

Then I found Groq. You make a key and start building. And the answers come back fast enough that you notice.

That combination, free and fast, is why it stuck. When you are a student, or you are testing whether an idea is worth anything, "free and fast" is the entire requirements document.

How to start in about two minutes#

1. Go to the Groq console and sign in.
2. Open API Keys and create a key.
3. Save it as an environment variable so it never ends up in your code:

export GROQ_API_KEY=your-key-here

4. Install the library and make your first request. This is the example from Groq's quickstart:

import os
from groq import Groq

client = Groq(api_key=os.environ.get("GROQ_API_KEY"))

chat = client.chat.completions.create(
    messages=[{"role": "user", "content": "Explain the importance of fast language models"}],
    model="llama-3.3-70b-versatile",
)

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

That is the whole setup. If you have used the OpenAI library before, the shape will look familiar, and that is on purpose: Groq's API is OpenAI-compatible, so a lot of existing code works by changing the base URL, the key and the model name.

What "free" actually means#

Free does not mean unlimited. Groq publishes its limits per model, and they are measured four ways: requests per minute, requests per day, tokens per minute and tokens per day. You hit whichever one you reach first.

Here are a few rows from the free plan table on Groq's rate limits page, as I read it on 12 November 2026:

These numbers change, and the list of models changes faster. Your exact limits are on the Limits page inside your own account, so check there before you plan around a number from a blog post, including this one.

A thousand requests a day is not enough to run a popular product. It is more than enough to build one, test it, demo it and show it to the first fifty people.

The things that will bite you#

• Limits are per organisation, not per key. Making five keys does not give you five times the quota.
• Tokens per minute is the one you hit first. Long prompts burn through 8,000 tokens quickly. Trim your system prompt and do not send the whole chat history every time.
• You get a 429 when you go over. The response includes a retry-after header telling you how many seconds to wait. Read it and wait that long instead of retrying in a tight loop.
• The models are open models. You are not getting Claude or GPT here. For a lot of jobs (classifying, summarising, extracting, quick chat) that is completely fine. For the hardest reasoning, test before you commit.

When to stop using the free tier#

When real people depend on it. The moment a 429 means a customer sees an error, move to the paid Developer plan or put a second provider behind it as a fallback. Free tiers are for learning and for proving the idea. They are a terrible thing to discover the limits of in production.

Who this is for#

If you are a student developer, or you want to build apps with AI and you have not paid anyone yet: try this before you do. Build the thing first. Find out if anyone wants it. Then decide who deserves your money.

This is not sponsored. I just really like the product, and I got tired of waiting for quotas to reset.