How I cut my AI inference bill by 80% - moving from OpenAI to self-hosted models