A free AI stack starts with the limits

Separate inference from search

Inference generates an answer. Web search retrieves material that an answer can cite. You may need only one. A summarizer for documents you already have can start without a search API; a research assistant may need both. Keeping that distinction explicit makes the prototype easier to test and its usage easier to understand.

Our current weekly finds compares Groq and Gemini within inference, and Tavily separately for web search. The order is an editorial starting point, not a measured model-quality ranking. Try a small set of representative prompts and record what fails before choosing a service.

Write down the boundary for each service

Groq offers a Free tier with limits that depend on the model and organization. Requests, tokens and audio can have separate limits. Its Developer tier introduces paid usage, so check the tier in your own console rather than borrowing a quota from someone else's screenshot.

Gemini lists free availability by model and feature. Check supported regions and the data-use conditions for unpaid services. A model appearing in the price table does not mean that every operation on it is free.

Tavily's Researcher plan lists 1,000 API credits per month without a credit card. Credits are usage units: basic and advanced search have different costs, and other operations have their own rules. Keep pay-as-you-go off if your budget is zero.

Make one narrow test

Choose ten non-confidential examples that resemble the task you want to support. For each, note whether the result is useful, whether sources support its claims, how much of the allowance it used, and how the app behaved on failure. Ten is a manageable test size, not a statistical guarantee.

Keep API keys on the server. Cap request size, limit retries and show a useful message when the provider rejects a request. A free prototype still needs a stopping point; an unlimited retry loop is not a recovery strategy.

Decide what you will keep

Keep the service if it solves the narrow task within your constraints. Change the design if the allowance is too small or the data conditions do not fit. You can also remove the feature: a curated source list may help your reader more than a generated answer you cannot verify.

Before sharing the prototype, repeat the test with the actual account tier and current documentation. Save the review date beside your notes. Free-plan terms can change, and an old comparison cannot promise tomorrow's access.