The Overnight

The Overnight — 2026-10-04

DeepSeek-Flash drops to $0.003 per million cached input tokens

What China's AI world did while the West was asleep. · 2026-10-04

Share: Share on X LinkedIn

The Number

  • DeepSeek-Flash now lists cache-hit input at $0.003/M off-peak, half that of peak, on its V4.1-Flash tier, per DeepSeek's pricing docs (2026-10-03).
  • DeepSeek-V4-Pro cache-miss input runs $0.66/M off-peak, 220x the Flash cache-hit rate inside the same house, per DeepSeek's pricing docs (2026-10-03).
  • Qwen3-max lists input at 2.5 yuan/M on the 0-32K tier, climbing to 7 yuan/M past 128K, per Alibaba DashScope pricing (2026-10-03).
  • GLM-flash-latest carries a $0.12/M prompt against a $4/M completion, a 33x input-to-output spread, on OpenRouter (2026-10-03).

Top of the Fold

  • Moonshot marks Kimi K3 as formally released at the top of its pricing docs, its new flagship tier now live, per Moonshot's platform docs (2026-10-03).
  • Otherwise a quiet morning up top; no marquee story crossed the wire overnight.

Untranslated

  • DeepSeek opens mass hiring for its elastic-compute team, calling specifically for senior engineers and posting a technical report as the job description, per 量子位 (2026-10-03).
  • DeepSeek retires the legacy deepseek-v4-flash and vision-exp model names, rerouting those requests to V4.1-Flash at Flash pricing, per DeepSeek's pricing docs (2026-10-03).
  • Otherwise quiet across the Chinese community overnight; no in-window exclusive cleared the gate.

The Backstory

  • Cloudwalk rival 云深处, the four-legged-robot maker named among Hangzhou's six dragons, swings back to a 8.8M-yuan first-half loss after a 2025 profit, R&D up over 140% as it ports its B-end playbook to humanoids, per its updated 科创板 prospectus via 虎嗅 (2026-10-03).

The Kicker

  • iQiyi's AIGC-made 灵魂摆渡 "浮生梦" short-film series crosses 10M yuan in revenue-share box office after its third installment dropped October 1, per 雷峰网 (2026-10-03).