At current rates of decline — approximately sixty per cent year-on-year across the major commercial providers — the cost of processing one million tokens will fall below three dollars before the close of 2026. For enterprise AI programme managers who built their business cases on 2025 pricing assumptions, this trajectory creates an unusual planning problem: the investments that were marginal at eight dollars per million tokens are now straightforwardly viable, and the applications that were rejected as economically impractical may warrant reconsideration.

The practical implication is not merely that AI is cheaper. It is that the architectural choices made under cost constraints — selective processing, aggressive chunking, retrieval-augmented rather than full-context approaches — may be suboptimal under the emerging cost regime. Organisations that designed their AI infrastructure for a cost environment that no longer exists should audit those design decisions.