Confirmed that mini uses ~30x more tokens than base gpt-4o using same image/same prompt: { completionTokens: 46, promptTokens: 14207, totalTokens: 14253 } vs. { completionTokens: 82, promptTokens: 465, totalTokens: 547 }.