The Number
- DeepSeek-V4.1-Flash lists at $0.006/MTok input peak on OpenRouter, the cheapest China-lab prompt rate on the board, per OpenRouter as of 2026-11-09.
- GLM-5.2 runs $6.8/MTok completion against a $0.0075 prompt rate on OpenRouter, the widest in-to-out spread among China labs listed, per OpenRouter as of 2026-11-09.
- Moonshot's Kimi K2 series stops onboarding new users from 2026-10-31 and sunsets entirely by late December, per Moonshot's pricing docs as of 2026-11-09.
Top of the Fold
- Alibaba ships Qwen Office 2.0 with a Team Agent built for humans and agents working the same doc, its first multi-person collaboration layer, per 量子位.
- Tencent releases Hunyuan 4 as an open model with a multimodal agent pairing, moving its open line up a tier, per 量子位 (company-stated; benchmark claims unverified).
- DeepSeek collapses its Flash lineup, retiring deepseek-v4-flash and routing old model names to DeepSeek-V4.1-Flash-1017 at Flash pricing, per its API docs as of 2026-11-09.
Untranslated
- 基元律动 (TokenRhythm) releases NeoHorse-1.5-Flash Preview, an agent-native model it frames as a second round of recursive self-improvement, per 量子位.
- WeChat Input Method quietly adds an AI voice recorder across macOS, iOS, Android and Windows with live Chinese-English transcription and structured summaries, per 智东西.
- 量子位 runs a first-person piece titled "AI ruined my CS degree," with students and a professor both leaning on Claude, a classroom-level take on model dependence circulating in China, per 量子位.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- 超聚变, the Zhengzhou server maker backed by Henan state capital, faces a 2026-11-12 Shenzhen listing review seeking 8B yuan, the largest ChiNext filing since the IPO slowdown, per 虎嗅.
The Kicker
- A "fill this empty room with what you know about me" prompt is sweeping Chinese feeds, after ChatGPT rebuilt one user's grey shell flat down to a silverfish he'd photographed a year earlier, per 新智元.
