The Number
- Moonshot's kimi-k3-max sits #2 on Arena's coding preference ranking at Elo 1675, the top China entry and behind only Anthropic's claude-opus-5-max, per LMArena as of 2026-08-01.
- DeepSeek's v4-flash cache-miss input runs $0.14/M, against $0.435/M for v4-pro, per DeepSeek's pricing docs as of 2026-08-02.
- Qwen's qwen3.7-flash lists at $0.03/M input on OpenRouter, the cheapest China-lab model on the board, per OpenRouter as of 2026-08-02.
Top of the Fold
- DeepSeek opens V4-Flash formal API to public beta, with company-stated agent scores it says beat V4-Pro-Preview, signaling the cheaper flash tier is now the push, per DeepSeek's API changelog.
- Moonshot lists Kimi K3 as its flagship at 1M-token context, moving the K-line up a full generation on its own pricing docs, per platform.moonshot.cn.
- ByteDance ships Seedance 2.5 doubling single-clip length, tightening the domestic video-gen race, per 极客公园.
Untranslated
- 长鑫 (CXMT) crosses a 4-trillion milestone cited in Chinese trade coverage, a scale marker the English press has not picked up, per 极客公园.
- WSJ-sourced Chinese coverage circulates that Coinbase and Airbnb swapped in Chinese models (Kimi K3, Qwen) to cut cost, framed in China as a repeat of the DeepSeek shock; circulating/unconfirmed, per 36氪.
- Zhejiang enacts a group standard defining the "AI OPC" one-person company, one core operator and under ten staff, a bureaucratic first that reads as folklore about solo AI shops, per 36氪.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- 虎嗅 reruns the tale of Leopold, the OpenAI super-alignment alum turned AI-accelerator fund manager whose leverage forced a liquidation, a 67% single-month drawdown told through his own apology letter to LPs, per 虎嗅.
The Kicker
- A Shanghai startup sends a "Teletubby" humanoid to clean apartments at 200 yuan/hour, billed straight-faced as pure manual intelligence, per 量子位.
