From OpenAI's first reasoning model to the next generation of models, AI inference costs have declined by two to three orders of magnitude within two years, while performance has improved significantly.