The Overnight

The Overnight — 2026-11-09

DeepSeek's V4.1-Flash cache-hit input drops to $0.006/MTok

What China's AI world did while the West was asleep. · 2026-11-09

Share: Share on X LinkedIn

The Number

  • DeepSeek's V4.1-Flash lists peak cache-hit input at $0.006/MTok and cache-miss at $0.3/MTok, with off-peak hours priced at half, per DeepSeek's docs as of 2026-11-08.
  • Qwen3.8-Max holds at 12 yuan per 1M input tokens and 36 yuan output on Alibaba's DashScope, unchanged across its Beijing, US, and Singapore deployments, per DashScope pricing as of 2026-11-08.
  • GLM-5.3 sits at $0.0699/MTok prompt and $1.4/MTok completion on OpenRouter, roughly 10x DeepSeek-Flash's input price among China-lab listings, on OpenRouter as of 2026-11-08.

Top of the Fold

  • Alibaba, Tencent, and ByteDance are all pushing AI office agents, with ByteDance's 豆包工作 drawing over 100,000 enterprise beta applications in its first week, a land grab more defensive than demanded, per 虎嗅.
  • Moonshot has formally released Kimi K3.1 Pro with a 1M-token context window and batch inference at 60% of standard price, while scheduling its entire K2 line for December retirement, per Moonshot's platform docs.
  • Otherwise a quiet morning up top.

Untranslated

  • 虎嗅 argues China's AI office boom is a market lacking commercial prospects, with the big platforms forced onto the table rather than choosing it, a dissenting read against the ad-saturated hype, per 虎嗅.
  • Embodied-AI training data has become the sector's costliest bottleneck, with industry-specific supply falling short of demand by about two orders of magnitude and no mature production model, per 财联社 via 虎嗅.
  • ByteDance spun its five-year internal AI-drug unit into Anew Labs, keeping 56% control after a reported $290M first round at a ~$1.5B valuation on a pre-clinical asset, per Z Finance via 虎嗅.
  • Capital is pulling out of "world models," per 量子位, which frames the retreat without naming the funds; circulating as a sentiment read, unconfirmed.

The Backstory

  • 中文在线 quietly sits behind both ReelShort and the new FlareFlow short-drama zone on Thailand's AIS carrier, where a VIP that costs $19.99 a week in the US is bundled at 39 baht a month, about 70x cheaper, per 刀客Doc via 虎嗅.

The Kicker

  • A Korean indie dev built a "reincarnation simulator" called 80亿分之一 with Claude in four hours, and by day three over 10,000 people were rerolling their next life at once as the link spread from Korea into Japan and China, per 酷玩实验室 via 虎嗅.