Which free LLM tier actually survived your production workload?

Every free tier is amazing until the day you actually depend on it: Groq hits RPM, Gemini throttles context, a provider silently degrades quality.

We ended up building a router that chains 25 free providers behind one OpenAI-compatible endpoint — circuit breaker, disk cache for repeated prompts, live dashboard of limits. It launched here on PH today as FreeGate (MIT, npm/Docker).

But before you reach for a tool: which free tier has genuinely survived real workloads for you, and for how long? Curious whether people run single-provider setups that just... work.