Tokens too cheap to meter

An AI model's price per token does not tell you what completing a job will cost. A cheaper model may generate more text to finish the same task, while a more capable model may finish with less work. The essay makes this distinction central to its argument that useful AI computation is becoming cheaper.

Its broader prediction is that inexpensive model decisions will become ordinary software components, including inside search and scheduling tools. That is a forecast about how developers might use falling task costs. Evaluating it requires comparing the cost of completing the same work at an acceptable quality, rather than comparing advertised text-generation prices alone.