I’ve been wondering about this since ChatGPT first came out. Tech bros have said inference costs have dropped by orders of magnitude, but I don’t think we can judge that by only what they’re charging.