The Verbosity Tax: Why Chinese AI's Benchmark Wins Come With a Hidden Cost
The week's model releases read like a flex: Alibaba's Qwen 3.8 Max announcement, Moonshot's Kimi K3 landing at #4 on Artificial Analysis's intelligence index with 2.8 trillion parameters, GLM 5.2 still riding high from its open-weight release. On raw benchmarks, the Chinese labs have arrived. Kimi K3 is priced at Claude Sonnet levels ($3/$15 per million input/output tokens) and, per Moonshot's cherry-picked figures, punches at Opus-tier performance. That is unambiguously good for the market. But the benchmark…