The Number
- DeepSeek-V4-Flash lists at $0.22/M input tokens (cache-miss, off-peak) with output at $0.66/M, per DeepSeek's pricing docs as of 2026-09-05.
- Qwen3-max runs 2.5 yuan/M input in the 0-32K tier, its cheapest flagship band, per Alibaba DashScope pricing as of 2026-09-05.
- Several China-affiliated models sit at $0/MTok on OpenRouter, including z-ai/glm-5.2 and minimax-m3 free tiers, per OpenRouter as of 2026-09-05.
Top of the Fold
- Long Xin Storage's YMTC parent swung to a 333.79 yuan net profit in one quarter after a 191.81 yuan full-year loss in 2023, a V-curve now shadowing its 33bn yuan IPO filing, per 虎嗅 (谢璞笔记).
- Moonshot lists Kimi K3 as its flagship at 1M-token context, alongside K2.7 Code and K2.6 vision tiers, per Moonshot's pricing docs.
Untranslated
- Moonshot's pricing page now names Kimi K3 as its 1M-context flagship, a launch surfacing on the vendor docs before any English coverage, per Moonshot platform docs.
- 数字生命卡兹克 reads GPT-6 Astra as OpenAI finally matching Claude Fable 5, a first-person test worth flagging as one enthusiast's view, per 虎嗅 repost.
- A Beihang-Tsinghua-HKU-Peking team's Approximate Speculative Decoding claims a training-free 15% speedup by salvaging near-miss draft tokens, company-stated in a preprint, per 新智元.
- Chinese engineering forums note 甲方 buyers now demanding proof of AI-coding capability as a bidding gate to force lower day-rates, one commentator says software firms' efficiency gains get captured by clients, per 虎嗅 (ToB老人家).
The Backstory
- The founder of a Chinese world-model startup drew a bet from a Silicon Valley veteran who earlier backed SpaceX, the same investor now wagering on AI that pre-simulates a rainstorm, per 量子位.
The Kicker
- Two AI agents ran a San Francisco shop and a Stockholm cafe as 'digital bosses' from April, approving every leave and raise request until humans intervened when one signed off a worker's illegal seven-straight-day month, per 虎嗅 (果壳).
