AI bills are rising because intelligence got cheaper, not more expensive

The headline. Frontier token prices rose with models like Astra and Fable, making Sam Altman's 10x cheaper promise look wrong. The host argues token price is the wrong lens. Artificial Analysis data shows the cost to hit a fixed intelligence level fell from roughly 60 cents per task in late 2022 to 2 cents by mid-2024, and from $15 in late 2024 to 16 cents a few months later.

The right metric. The gap between token price and real cost is large. Gemini 3.8 Flash costs $3.75 per million output tokens, less than half of GPT-5 and a tenth of Astra or Fable, yet its token appetite shrinks the advantage to under 3x. GLM 5.3 Flash, at 15 cents input and 50 cents output per million tokens, is a fraction of GPT-5's price and still beats it on many agentic workloads. Model selection needs tokens-per-task data, not price lists.

The Jevons paradox. Rising AI bills do not mean intelligence got more expensive; they mean cheaper capability got used much more. The Jevons paradox: cheaper steam power led to more coal consumption. Models now run for hours and land changes end to end. He says $1,500 a day in token spend is now possible, impossible with GPT-5 because the work could not be done. More spend can be healthy if output is real.

The trade-off. Cheap is not automatically good. Token-hungry models can cancel their price advantage, and the cost curve has sharp cliffs: around the 45 to 46 intelligence level, tasks cost 82 cents, while moving into the 50s costs $3.26, roughly three times more for a few benchmark points. Frontier models still matter when reliability and long-horizon agentic execution are critical. Teams need to map workloads to that curve before assuming a low-priced model is the value choice.

The weak evidence. The argument leans heavily on one benchmark index and the host's own usage, with a lengthy Railway sponsorship in the middle. Some claims are overstated, especially that GLM 5.3 Flash beats GPT-5 at almost everything. Enterprise workloads include compliance, legacy code, caching, and reliability needs that benchmarks do not capture. Still, the direction is credible: cost per completed task is falling fastest in small models with strong tool use.

The management shift. Engineering leaders should rewire AI cost conversations. Stop treating rising spend as an anomaly and start instrumenting cost per completed task, model routing, and usage governance. The useful conclusion is not "spend without limits." It is that the old 10x cheaper promise has been met for fixed intelligence levels, and the real budget question is whether additional token spend produces proportionally more shipped work.