Everyone was busy watching the GPT-vs-Gemini title fight when a phone company from Beijing quietly walked into the ring, threw one punch, and took the belt. Xiaomi. The people who sell you rice cookers. That Xiaomi.

The Model Nobody Was Betting On Wins the Volume Game

Xiaomi's MiMo-V2-Pro is now the single most-used model on OpenRouter by weekly token volume, processing 4.65 trillion tokens for a 22.3% platform share — more than double the runner-up, and comfortably ahead of OpenAI's entire lineup. It's a trillion-parameter mixture-of-experts model with only 42 billion active parameters per request, a million-token context window, and pricing of $1/$3 per million input/output tokens against Claude Sonnet 4.6's $3/$15.

Performance isn't an afterthought bolted onto the cheap price tag, either: 78% on SWE-bench Verified versus Opus 4.6's 80.8%, and eighth globally on the Artificial Analysis Intelligence Index. It got so convincing during its stealth "Hunter Alpha" testing phase that people mistook it for DeepSeek's next release.

Cheap, Fast, and Suspiciously Good at Coding

This is the part that should worry the incumbents: developers don't pick models out of brand loyalty, they pick them because the invoice is smaller and the code still works. When a model lands within a few points of the frontier for a fraction of the cost, "good enough" starts looking like "obviously correct choice," and token volume follows the spreadsheet, not the marketing deck.

The quieter signal is what this says about the AI price war shifting east. While the loudest headlines fixate on which US lab has the smartest model, actual usage is voting with its wallet for whoever ships something 78%-as-good at 20%-of-the-price.

Somewhere in Cupertino and San Francisco, someone's updating a competitive analysis slide, and it is not a fun slide to build.

Source: Decrypt