KOST · $/M tokens
$–
The price of a thought. What does the world actually pay per million tokens?
A volume-weighted blend of the realized price of LLM inference across every major model — built from OpenRouter's ~20 trillion tokens a week (price × real token volume). The open, transparent analog of the token-cost benchmark otherwise locked behind Bloomberg.
The blended price the market pays per million tokens, since September 2025. It falls as traffic rotates toward cheap open models, and rises as demand leans back into frontier models.
Each model's blended price is weighted by its actual daily token volume on OpenRouter — not an equal-weight basket. A $0.10 open model carrying 10% of traffic counts ten times a frontier model carrying 1%. This is the realized cost-to-serve, not a list-price average.
Each model's price blends input and output rates at an assumed 80% input share. We publish the full sensitivity band openly — the benchmarks that charge for this number don't.
Pre-launch history holds each model's price at today's level and varies only the usage mix — capturing the dominant driver (demand rotating between cheap and frontier models). From launch forward, every tick records true price × volume.
KOST is published as a reference index a contract can settle against. Each month settles to the arithmetic mean of the daily index over that calendar month — Asian-style, the same mechanism the new GPU-compute futures use — finalized on the 1st. Daily fixings, immutable once final.
Volume-weighted, transparent, and free — the price of intelligence in one number.