Why it matters
The most-routed model on OpenRouter may not be a US lab's flagship but an anonymized Chinese MoE tuned for cost per token. When buyers route by price and context, an MIT-licensed model firing only ~48B of 1.6T parameters reshapes where agent spend lands.
Tokenmaxxing read
MoE is a tokenmaxxing lever: activating ~48B of 1.6T parameters cuts cost per token, and a 1M-token window feeds whole repos to agents. WinBuzzer flags the headline 10.1T-monthly / 559B-daily OpenRouter figures as unverified — the leaderboard signal is real, the volume isn't.
Source takeaway
Markus Kasanmascheff (WinBuzzer, June 30 2026): LongCat-2.0 is MIT-licensed, trained on 35T+ tokens on application-specific 'AI ASIC superpods' with no chip vendor named, and lands as GLM 5.2 and DeepSeek price cuts squeeze coding-model economics.
