The study shows that per-token list price is an unreliable proxy for actual LLM operating cost, with thinking-token variance and run-to-run randomness making true COGS unpredictable — a direct risk for any product built on flat-fee pricing over variable model usage.