The Number
- DeepSeek's cheapest listing, deepseek-v4-flash, runs $0.0023/MTok input, the floor of China-lab pricing on OpenRouter as of 2026-10-06.
- DeepSeek-flash prices input at $0.15/MTok cache-miss off-peak against $0.66 for v4-pro, per its official docs on 2026-10-06.
- Qwen3-max thinking input holds at ¥2.5/M under 32K tokens, the entry tier on DashScope as of 2026-10-06.
Top of the Fold
- DeepSeek is closing a new round of at least ¥80B with Tencent and CATL among the largest backers, paving a 2027 IPO after demand outran its ~¥50B target, per 智东西 citing Bloomberg.
- Moonshot has wrapped its final private round at a reported ~$50B valuation, pricing the Kimi maker alongside DeepSeek as the K3 flagship ships, per 智东西 citing Bloomberg.
Untranslated
- Moonshot has quietly listed its K3 flagship with a tiered cache-write price split into 5-minute and 1-hour TTLs, a billing structure no rival exposes, per its pricing docs on 2026-10-06.
- DeepSeek has retired the legacy deepseek-v4-flash and deepseek-v4-flash-vision names, routing those calls to DeepSeek-V4.1-Flash at the flash price, per its API docs on 2026-10-06.
- Otherwise quiet across the Chinese community overnight; the wires ran mostly on OpenAI and Nobel items outside the lane.
The Backstory
- Xiaomi's new 澎程 extended-range line cleared 10,000 deliveries in its first month, breaking in while H1 extended-range wholesale fell 13.1%, per 极客公园.
The Kicker
- 量子位 ran a mock-alarmed post asking whether Terence Tao has joined the AI slowdown camp, the kind of headline the Chinese math-AI crowd forwards for the pun, per 量子位.
