Theo Browne distinguishes price per token, tokens consumed per task and the cost of reaching a given capability level. Against Sam Altman’s prediction of cheaper intelligence, he argues that comparing today’s premium model with an older premium release misses much cheaper options that can now handle similar work.
Theo Browne uses benchmark and pricing comparisons to show why token prices alone are misleading: a model with inexpensive tokens can consume enough of them to narrow its cost advantage. He emphasizes that input-heavy classification, cached agentic coding and long-running tool use have different cost profiles, so the relevant comparison depends on the workload.
Theo Browne attributes increasing total spend partly to models doing longer, harder end-to-end tasks that previously required more human intervention. He relates that demand expansion to Jevons paradox and warns that changing, saturated benchmarks make capability comparisons imperfect. His forecast of continued cost reductions is an argument about the trend, not a guarantee of any particular future model’s price.
Watch on YouTube
