I had Gemini train its own replacement for $9

A developer used Google's Gemini API to label 4,290 Reddit comments about chef's knives for $9, then fine-tuned an open-source named-entity recognition model (GLiNER) on those labels to achieve 0.83 F1 accuracy—matching Gemini's performance at near-zero cost for future predictions. The trained model pays for itself after the initial 4,290 comments, making it far more economical than continued API calls, though the model was evaluated against Gemini's labels rather than human-verified ground truth.