Free-tier LLM gateways: test the fallback before you commit

A free tier is a trial allowance, not a permanent price. The useful question is what happens when that allowance runs out: does the request stop, switch to a paid model, or fail silently?

Test the boundary, not the homepage

Start with a task small enough to finish inside the free allowance, then watch the response when the allowance is exhausted. A gateway that falls through to paid capacity should make that visible, and a gateway with no balance should return a clear error instead of a timeout.

Check the model list first

The model list should be readable before signup. If the list is behind a login, you cannot verify which models are callable or which rows are marked free. A public endpoint such as is the quickest evidence.

Check compatibility

If the gateway exposes an OpenAI- or Anthropic-compatible base URL, switching clients is a configuration change rather than a rewrite. That matters more than a long feature list.

Set a budget

Before the first paid call, set a budget in the console. A fallback that runs without a ceiling is not a safety net.

Summary

Pick a gateway by testing the failure path: exhausted free quota, provider error, and an incompatible client. The successful demo is the easy part.