84% of tokens processed on OpenRouter are not state of the art. Six models carry 80% of that volume, deliver 77% of frontier performance, and cost $0.50 per million tokens. Claude Fable 5, the current SOTA benchmark, runs $20 per million. That 40x price gap is not a rounding error. It is a market signal.
The frontier itself moves fast: models are two-thirds smarter than they were last November, with two new releases every three days and 3-5 Artificial Analysis point jumps landing roughly each quarter. But Ramp's spend data shows buyers are price-elastic. Fable 5 captured only 6% of Anthropic token volume and 11% of spend a month after launch. GPT-5.6 Sol held about a quarter of OpenAI tokens. The best open-weight model hit 80% of the frontier score by May 2026, up from 48% a year prior. Share is not following performance.
The argument worth reading in full is structural, not tactical. If SOTA share stops growing with each new release, a nine-figure training run has to recoup cost against a shrinking addressable market. The original piece names the exceptions where frontier models earn their price: software engineering, security architecture. It also tracks how portfolio companies are defaulting to fine-tuned and open-source models optimized for price over performance. The frontier that shapes most deployment decisions is not the intelligence index. It is the cost curve below it.
[READ ORIGINAL →]