Those free-tier caps match what I’ve been hitting on Streamlit prototypes too. Once Gemini/Groq/OpenRouter free windows are gone for the day, I usually drop the app onto a Haiku-class model instead of keeping a Sonnet/Opus key in .env for every refresh. On Anthropic’s public list that’s about $0.10 per 1M input / $0.50 per 1M output for Haiku 5.5, versus roughly $2/$10 for Sonnet 5.5.