Benchmarks lie; production traffic tells the story. The real cost driver was reprocessing, not inference. Two cents after running this through prod traffic.