84% of AI tokens avoid state-of-the-art models; cheaper alternatives capture majority

The article presents a striking divergence between AI capability progress and actual buyer behavior. State-of-the-art models have improved significantly—two-thirds smarter than last November—and the pace of release remains intense at two new models every three days. Yet this frontier innovation is not translating into dominant market adoption. Instead, 84% of tokens on OpenRouter route to non-frontier models, a gap that reflects rational economic decision-making rather than capability gaps.

The core dynamic is price elasticity. Six cheaper models generate 77% of frontier performance at 2.5% of Claude Fable 5's cost, and actual spending patterns confirm buyers respond to this trade-off. Fable 5, despite launching with frontier capabilities, captured only 6% and 11% of Anthropic's tokens and spend respectively a month after release, while generating less total model-attributed revenue than a pricier OpenAI tier. This pattern repeats across vendors: buyers consolidate on models that clear their workloads efficiently, not on capability alone. The best open-weight model's improvement from 48% to 80% of frontier performance in a year compounds this pressure—the performance gap that justifies premium pricing is shrinking.

The article identifies two distinct Pareto frontiers. Frontier models retain clear value for software architecture and security design, where the best available option justifies its cost. But application deployment optimizes differently: toward cheaper, fine-tuned, or open-source alternatives. If this second frontier becomes the dominant use case—and the data suggests it is—then the venture economics of frontier training runs face a rising bar. A nine-figure investment must recoup its cost through market share, yet each new state-of-the-art release should capture less share than the last, because the alternative already performs well enough.