Open-weight share
–
Who is actually winning inference? Open vs closed, US vs China — settled by where the tokens go.
The live structure of the LLM market, read straight from OpenRouter's token routing across ~340 models. Not press releases, not benchmarks — whose models the world actually runs, measured every day. Indices no one else publishes.
Every lab's share of routed tokens since September 2025. Warm bands are Chinese labs, cool bands US — watch the colour of the chart flip as open-weight models took over.
Share of routed tokens today, with the change over the last week.
OpenRouter clears ~20 trillion tokens a week across hundreds of models and reports each model's daily token count. We map every model to its lab, country, and open/closed status, then sum the shares. It's revealed demand, not stated intent.
Open-weight share = tokens on downloadable-weight models. China share = tokens on Chinese-lab models. The Herfindahl index (–) tracks whether demand is consolidating into a few labs or fragmenting across many.
The volume-weighted price of closed frontier models divided by the open-weight average — how much extra the frontier still commands. Pairs with KOST, the blended cost-of-inference index.
The open-weight share is published as OPEN — a reference index a contract can settle against. "Will open models clear X% of inference by quarter-end?" has real conviction behind it and no instrument; OPEN is the settlement oracle. Each month settles to the mean of the daily share (Asian-style), finalized on the 1st.
This monitor is the overview. Every index below is also published as a standalone, settleable contract.
Open vs closed, US vs China — live, transparent, and free.