The Number
- DeepSeek lists deepseek-flash cache-hit input at $0.003/M tokens off-peak, with cache-miss input at $0.15/M, per its pricing docs on 2026-10-05.
- GLM-5.2 runs at $0.0192/M input against $16/M completion on OpenRouter as of 2026-10-05, the steepest input-to-output spread among China models listed.
- Qwen3-max sits at 2.5元/M input in its lowest 0-32K tier, climbing to 7元/M past 128K, per Alibaba DashScope pricing on 2026-10-05.
Top of the Fold
- Moonshot marks Kimi K3 as formally released on its platform, moving its flagship into a new tiered-cache pricing band, per its official pricing docs.
- DeepSeek retires the legacy deepseek-v4-flash names and routes those calls to DeepSeek-V4.1-Flash at Flash prices, folding older variants into one model, per its API docs.
Untranslated
- 魔芯科技 ships 摸小宠, a HarmonyOS-exclusive app that turns a few cat photos into a walk-around 4D digital pet, debuting on Huawei's Mate 90 line, per 新智元 (company-stated capability, verify).
- Otherwise quiet across the Chinese community overnight: wires ran mostly on the Nobel medicine prize and overseas labs, with no China-native model drop passing the gate.
The Backstory
- 虎嗅 maps the world-model land grab back home after AMD's $8.2B all-stock buy of World Labs, asking which Chinese entrants get absorbed and which break out, per 虎嗅.
The Kicker
- 极客公园 takes apart why AI image tools keep generating the same pretty women, tracing it to a fake 'ordinary Korean woman' clip that hit 15.25M views before fans spotted a retired batter still on the scoreboard, per 极客公园.
