Open-weight models now generate 69.1% of named token volume on OpenRouter, with closed models at 30.9%. This is not the whole AI economy, but OpenRouter sits at the API frontier where developers switch models daily and route each request to the best available option. The shift has been sharp and it has happened since the start of 2025.
The leadership board inside open models is moving fast. DeepSeek held an early lead, then MiniMax and Kimi took share in late 2025 and early 2026. Launches from MiMo, Qwen, Tencent's Hy3, and DeepSeek again reshuffled the rankings. US lab Arcee has made a strong recent appearance. Each clustered wave of releases triggers large-scale developer testing and then locks in a new, higher plateau of sustained token volume.
Open models still represent a fraction of total inference compute globally. That caveat matters. What also matters is the pattern: production traffic is following experimentation. The original piece is worth reading for the token-volume charts, which show exactly how each model launch creates a step-change in usage rather than a gradual climb. The question it leaves open is which of these challengers holds share long enough to matter.
[READ ORIGINAL →]