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

The article begins with a paradox: state-of-the-art AI models are advancing faster and becoming much smarter than they were in November, with labs shipping two new models every three days. Yet despite this frenetic pace of innovation, 84% of tokens on OpenRouter—a popular aggregator platform—are not routed to state-of-the-art models at all.

This divergence reflects a hard economic reality. Six models that are not state-of-the-art collectively account for the supermajority of token volume, delivering approximately 77% of frontier performance while costing just 2.5% of what Claude Fable 5 costs per token. In the week of August 10, these six models had a blended price of $0.50 per million tokens, compared to Fable 5 at $20. For most applications, this price-to-performance trade-off favors the cheaper options decisively.

Buyer behavior confirms this pattern. Fable 5, which launched at a frontier capability level, captured only 6% of Anthropic tokens and 11% of Anthropic spend a month after launch. Despite being substantially more expensive, it generated roughly 75% as much model-attributed revenue as GPT-5.6 Sol, OpenAI's priciest mainline tier. This reveals strong price elasticity: buyers will trade frontier capability for dramatically lower cost if the remaining performance is sufficient for their task.

The gap between frontier and alternatives is also tightening from below. The best open-weight model reached 80% of frontier performance by May, up from 48% a year earlier. As performance thresholds for common workloads are met by cheaper alternatives, the justification for paying a premium erodes.

The article identifies two different contexts where frontier models retain clear value: software engineering architecture and security design, where the best available model's superior reasoning and robustness justify its cost. But for application deployment, teams increasingly default to smaller, fine-tuned, or open-source models optimized against a different Pareto frontier—price over raw performance.

If this pattern persists, the economics of frontier model training become untenable. Each new state-of-the-art release will capture less incremental market share than the one before, because the prior frontier model (now standard or open-source) already performs well enough. When share stops shifting and "good enough" truly becomes sufficient, a nine-figure training investment will struggle to recoup its cost—especially as the bar for what counts as "good enough" rises with time and competition from cheaper alternatives.