29/07/2026
Token prices collapsed. Your AI bill did not.
TechNewsWorld reported yesterday that token costs have fallen roughly 98% since early 2024. That is a real number from Windsor Drake's Jeff Barrington, verified on the page this morning.
Enterprise AI spend kept climbing anyway.
73% of enterprises blew past their original AI budgets last year. The reason is not compute cost. It is consumption shape.
A chatbot answers one prompt. An agent fans that prompt out into planning, tool calls, retrieval and verification. One instruction becomes hundreds of calls and five to thirty times the tokens. Actual output in agentic workloads is often only 5% to 15% of tokens consumed. The rest is context overhead.
PDFs are a perfect example. They were designed for rendering, not understanding, so every PDF forces the model to burn tokens reconstructing layout before extracting facts. You are paying a reasoning model to do a formatting job.
Inference is now around 85% of enterprise AI budgets. The meter is on.
Cheaper per-unit cost does not mean cheaper total cost if the architecture wastes it. Same problem, different domain.