The Number
- DeepSeek lists deepseek-v4-flash cache-hit input at $0.007/M off-peak, half its peak rate, per its pricing docs (2026-08-30).
- Qwen3-max sits at RMB 2.5/M input under 32K tokens, its cheapest flagship tier, per Alibaba DashScope pricing (2026-08-30).
- Zhipu, MiniMax's M3 and MiniMax M2.7 all list at $0/M prompt and completion on their free tiers, on OpenRouter (2026-08-30).
Top of the Fold
- Qualcomm says its IQ10 robotics platform has shipped to multiple customers and robotics products are entering mass production in China, positioning its physical-AI stack against Nvidia on the device side, per 智东西.
- OpenAI cuts off Cursor's direct model access effective November 12, citing a trust breach after SpaceX's acquisition, closing a channel Chinese coding-tool builders had watched, per 极客公园.
Untranslated
- Kimi's platform now flags Kimi K3 as the shipped flagship at 1M-token context, with the legacy Moonshot V1 line scheduled to go dark across all platforms August 31, per Moonshot's pricing docs.
- DeepSeek quietly lists a deepseek-v4-flash-vision-exp variant at cache-miss $0.22/M input, its first vision-capable Flash tier, per DeepSeek's pricing docs (2026-08-30); a listing, not a capability claim.
- Huxiu carries a widely-shared essay arguing AI is repricing effort itself, with judgment becoming the scarce paid skill while diligence deflates, framed as one contrarian view, per 虎嗅; circulating commentary, not reporting.
- A two-person Chinese team's fully AI-generated live-action short posing as an American teen drama has drawn million-level engagement on Chinese platforms with no US actors, sets, or director, per 虎嗅.
The Backstory
- Software companies are now racing into hardware bodies for their models: ByteDance's TRAE co-branded with FoloToy to build a wearable digital badge, as ambient AI hardware prices slide roughly $40 off the prior generation, per 极客公园.
The Kicker
- Fields medalist Terence Tao posts a paper reporting that four frontier AI systems produced near-flawless solutions to 7 of 10 unpublished research-level math problems for tens to hundreds of dollars each, per 新智元; preprint, per the linked arXiv report.
