I Tried 5 Free LLM APIs With One Python Script

I conducted tests on five free Large Language Model (LLM) APIs using a single Python script for small tools and experiments. The providers were Groq, Google AI Studio, OpenRouter, Cloudflare Workers AI, and Token Harbor. I did not add a credit card to any account and used the same Python script and tasks as much as possible. The tests were conducted on August 28–29, 2026.

For Groq, I used the qwen/qwen3.8-27b model with limits of 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute, and 2,000,000 tokens per day. The coding task completed in 1.85 seconds, and the context test passed in 1.85 seconds. The model found all three values in the synthetic incident log.

Google AI Studio required a few minutes to set up but showed 15 requests per minute, 250K input tokens per minute, and 500 requests per day. The coding task took 6.54 seconds to complete, but one request failed with a 503 error before retrying. The context test passed after retrying after about one minute, taking 24.35 seconds. The model found all three values in the synthetic incident log.

OpenRouter also did not require a credit card or further email verification. The model poolside/laguna-s-2.1:free had a rolling 7-day value allowance. The coding task failed initially with a 429 error, but a manual retry succeeded. The context test failed twice with 429 errors, but manual retries succeeded. The model found all three values in the synthetic incident log after retries.

Cloudflare Workers AI had a longer setup process but showed 10,000 neurons per day. The coding task completed in 5.93 seconds, but the model had one edge-case bug in the returned code. The context test passed in 5.93 seconds after reducing the input size. The model found all three values in the synthetic incident log.

Token Harbor required no credit card information and showed a rolling 7-day value allowance. The coding task completed in 1.84 seconds with 63,887 input tokens. The context test passed in 34.62 seconds with 63,887 input tokens. The model found all three values in the synthetic incident log.