Xiaomi's MiMo model family has topped OpenRouter's usage rankings, drawing roughly 22% of all traffic on the model-routing platform, about three times OpenAI's share, per OpenRouter's rankings data. The company has now open-sourced MiMo-V2.5-Pro, a one-trillion-parameter mixture-of-experts model with frontier-level coding performance, available on Hugging Face under a permissive license. A phone maker has become one of the most-used AI model providers by developer token volume, and price is the mechanism.
The number that frames the shift is that 22% of OpenRouter traffic against OpenAI's roughly 7.5%. OpenRouter is where developers comparison-shop models by cost and capability per token, so it is a cleaner read on what builders actually deploy than any lab's announced user count. Xiaomi's models win there because they deliver competitive coding benchmarks at a fraction of the per-token price, and because open weights let teams self-host and avoid API costs entirely. When the model is good enough and the weights are free, the US labs' pricing power on commodity coding tasks erodes from underneath. MiMo-V2-Pro even topped the charts anonymously as "Hunter Alpha" before Xiaomi claimed it, which tells you developers picked it on merit, not brand.
The strategic read for the US frontier labs is that their API revenue thesis faces a floor problem. The valuations attached to OpenAI and Anthropic assume durable pricing power on inference. Open-weight models from well-capitalized Chinese firms, given away to win developer mindshare and hardware ecosystems, put a ceiling on what the labs can charge for anything short of frontier-only capability. Anthropic's Sonnet 5 price cut this week is a response to exactly this pressure. The labs keep the frontier; the commodity middle is being competed to near-zero.
Bottom Line
Developers are voting with tokens, and a growing share goes to open-weight models that cost a fraction of the US labs' APIs. For anyone modeling OpenAI or Anthropic revenue, the open-weight price floor is now a real input, not a hypothetical.