If you use AI for rubber ducking, simple trick to reduce your costs.

Use a cheap model and let rip, I have been using qwen 235b for this. Then periodically switch to a higher quality model like Gemini 2.5 pro and ask 1 "please review this conversation and add any notes or insights you have" query. Then back to the cheap model for a while.

Easy way to get a lot more work done without burning through all your credits and without sacrificing accuracy in the end.