“But inference has gotten cheaper even faster than models have gotten better, and nobody has an intuition for something becoming 1000x cheaper in two years.”
A Gemini Flash query costs about a twenty-fifth of a Bing search API call. DeepSeek’s published numbers show roughly 80% margins on inference. OpenAI’s losses come from not monetizing free users, not from serving them. The cost lands on websites that agents scrape for free.