5 LLM Cost Mistakes I Made (And the One Fix That Saved 40%)

Sharing my actual mistakes so you can skip them. Mistake 1: Using Sonnet for File Reads Paying $0.015/1K tokens to read a config file. DeepSeek does this for $0.0014. Identical output. Mistake 2: Long Sessions By message 30, every prompt carries 80K+ tokens of history. I was paying for context I did not need. Now I start fresh sessions aggressively. Mistake 3: Not Tracking Per-Task Costs Had no idea where my money went until I logged every API call for a month. 60% was routine tasks at premium prices. Mistake 4: Manual Model Switching Tried switching models manually for a week.