GLM-5.2 Adoption Tracks 10-Point Jump in Chinese AI Traffic Share

The Atlantic piece records praise from Andreessen, Rauch, and unnamed founders but stops at cost comparisons. Primary data show the model undercuts Anthropic and OpenAI agent pricing by multiples while matching coding benchmarks, triggering direct downloads of its open weights. Ramp and OpenRouter metrics capture only tracked spend; self-hosted instances remain invisible yet scale faster among developers avoiding per-token bills. U.S. labs face margin pressure after locking in corporate pilots at premium rates. Uber exhausted its 2026 Anthropic allocation early; Citi restricted access to frontier models. These constraints arise from training and inference costs that Chinese labs offset through state compute subsidies and lower labor overhead, an incentive structure the article notes only in passing. The pattern repeats DeepSeek's January 2025 launch, when RAND recorded Chinese model traffic rising from 3 to 13 percent globally within two months. GLM-5.2 extends that shift into agent workflows, where coding and task automation now carry measurable labor displacement risk rather than chat-only usage. Forward indicators point to accelerated price compression. If U.S. firms replicate the 2025 response by releasing cheaper agents, adoption curves will still favor the lower baseline set by Chinese open models, widening the gap in accessible tooling for non-enterprise users.