The Number
- DeepSeek-V4-Pro runs $0.66/M input on cache-miss peak, against Qwen3.8-max at 12 yuan/M, per DeepSeek's pricing docs as of 2026-08-22.
- Kimi K3 Max sits #2 on Arena's coding preference ranking at 1674 Elo, behind only Claude Opus 5 Max, per LMArena as of 2026-08-22.
- DeepSeek moves off-peak rates to all weekend hours starting 00:00 Beijing time Sunday 2026-08-23, per DeepSeek's pricing docs.
Top of the Fold
- DeepSeek ships V4-Flash-Vision-Exp on its API, adding image input to a line that until now was text-only, per its official channel via 极客公园.
- Qwen3.8-max lands #2 on Arena's vision preference ranking at 1302 Elo, the only non-Anthropic model in the top three, per LMArena as of 2026-08-22.
- Pony.ai reports Robotaxi revenue up 534% to $20.6M in H1 2026 while net loss widens 9.1%, the growth-without-profit split 虎嗅 flags as the open question, per 虎嗅.
Untranslated
- Kuaishou's Kling AI clears 850M yuan in Q2 revenue, up over 200% year on year with 100M cumulative users, the one bright line in a quarter where group revenue grew 1.4%, per 虎嗅.
- Beijing launches a 1.6M-square-meter AI-for-Science cluster in Haidian's Xisanqi, alongside the first national standard for autonomous labs, per 智东西.
- UBTech runs a customer production line 1:1 inside WRC 2026, pitching one embodied brain across industrial, commercial and home robots, per 量子位.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- An Economist study of 27,000 Chinese middle-schoolers finds AI users' homework scores rose 18% while their closed-book exam scores ran 20% below non-users, as recorded by 极客公园.
The Kicker
- At WRC 2026 the booths stopped competing on dancing and pouring drinks and started competing on whether a robot can work a steady eight-hour shift, per 雷峰网.
