When I started building Xandhi OS - an AI-native app builder - every advisor and Twitter reply told me the same thing: "Just use GPT-4. Stop overthinking it." I didn't. Here's what happened, with real observations, real failure modes, and zero marketing varnish. The thesis The thesis was simple: for code generation in 2025, the gap between top free models and GPT-4 has collapsed for most tasks - and where it hasn't, you can route around it. If that's true, building on free-first models means: Dramatically lower cost per build Permanent free tier for users (real competitive advantage) No vendor lock-in to any single provider's pricing or roadmap If it's wrong, I quietly migrate to GPT-4 and eat the cost. So I tested. The contenders Through OpenRouter, I had access to dozens of models.…