A lot of "free" AI tools quietly mean "free, but on the oldest and slowest model we have." Lumen's Free tier works differently: instead of committing to a single provider, a free question is answered by whichever of several fast, efficient models is actually up and responding quickly right now -- with automatic retry and backoff if a provider is slow or returns an error, so one bad upstream response doesn't become your problem.
In build mode the cascade goes further: a request can walk through several free-tier legs (including a fast open model served over NVIDIA's inference API) before ever touching a paid model, so free users still get a real, current model behind their answer rather than a token gesture.
Pro removes the cascade entirely and lets you pick a named frontier model directly -- Opus, GPT-4o, Grok or Gemini -- for when you specifically want that model's behavior rather than whichever free leg answered fastest. See the current lineup on Pricing.