Like other posters said, maybe a local model is a good option. I've found the Qwen3:4B (reasoning) model works pretty well for many things.

I'm planning to run a local model on a $149 mini-pc and host it for the world from my bedroom. You can read a bit more about my thinking below.

https://joeldare.com/my_plan_to_build_an_ai_chat_bot_in_my_b...

These hosted models are better but it feels like the gap is closing and I hope it continues to close.