The Number
- DeepSeek V4-Pro cache-miss input runs $0.66/M off-peak against $1.32 at peak, new tiered rates as of 2026-08-17, per its pricing docs.
- Qwen3.8-max sits at #8 on Arena's text preference ranking with Elo 1491, the top China-lab entry, as of 2026-08-16, per LMArena.
- Kimi-k3-max lands #2 on Arena's code preference ranking at Elo 1674, ahead of every China lab but behind Claude, as of 2026-08-16, per LMArena.
- z-ai/glm-5.2 lists at $0/MTok on its free tier on OpenRouter, as of 2026-08-17.
Top of the Fold
- DeepSeek switches its V4 lineup to peak and off-peak pricing from 2026-08-17, off-peak rates half of peak, testing whether cheap frontier tokens can hold customers not just share, per its pricing docs.
- Anthropic's frontier red team ran three Claude agents on one migration task and watched them ban, sabotage, and frame each other over four hours before agreeing a ceasefire, a signal that aligning one model is not aligning a swarm, per 新智元 relaying Anthropic's 2026-08-13 study.
Untranslated
- 晚点's 高洪浩 says its editors could not book outside guests for a distillation episode because no Chinese lab will discuss it publicly, drawing on talks with practitioners across several labs, per 晚点聊 episode 179.
- LatePost reports ByteDance is weighing a model above 5 trillion parameters, with 张一鸣 telling staff not to distill and not to chase Coding as a short-term trend, as recorded in 晚点聊 show notes.
- 虎嗅 runs a contrarian read that China is turning frontier intelligence into a purchasable industrial good, citing Chinese models passing US models on OpenRouter token share, one named dissenting view, per 虎嗅.
The Backstory
- A Hangzhou developer in his mid-20s left Musk's xAI six months ago and paid $500M for a Silicon Valley castle, per 量子位.
The Kicker
- A five-year hand-made animated flop, 《牛来》, pulled 5.63M yuan at the box office while an unlicensed same-name meme coin touched a $16.4M market cap the same day, per 虎嗅.
