The Number
- DeepSeek-flash lists input at $0.003/M tokens on a cache hit off-peak, against $0.3/M on a cache miss at peak, per DeepSeek's pricing docs on 2026-10-04.
- Zhipu's GLM-5.2 shows an $8/M-token completion price against $0.05/M on prompts, roughly a 160x gap, on OpenRouter as of 2026-10-04.
- Alibaba's top-tier Qwen3.8-max-prime runs ¥24/M input tokens, the priciest rung in the Qwen line, per DashScope on 2026-10-04.
Top of the Fold
- Moonshot's Kimi K3 flagship line goes live on its pricing page with per-TTL (5min/1h) cache billing, pushing its context-cache economics harder, per Moonshot docs on 2026-10-04.
- Otherwise a quiet morning up top; the wires ran chip-supply and US-lab stories outside our lane.
Untranslated
- Chinese labs are rushing into 'decision models,' with Shanghai AI Lab open-sourcing Intern-Decision and 陈大年's five-month-old StartLux claiming (company-stated) its StartLux-Decision beats TypeSafe's Jev, per 极客公园 on 2026-10-04.
- 'Forward-deployed engineer' is being written up as the hottest AI hire in China at a reported ¥50,000/month, per 量子位; treat the figure as wire-reported, not surveyed.
- Otherwise quiet across the Chinese community overnight.
The Kicker
- Job seekers' griping about the 'uncanny valley' of AI-run interviews climbed China's trending-search list overnight, per 极客公园.
