The Number
- DeepSeek's new deepseek-flash lists at $0.15/M input (cache-miss, off-peak) with a 1M-token context, per its pricing docs as of 2026-10-18.
- DeepSeek-V4-Pro runs $0.66/M input off-peak against $0.0052/M for the cheapest V4-flash route on OpenRouter, per OpenRouter as of 2026-10-18.
- Qwen3.8-Max holds at ¥12/M input standard tier, with Batch calls billed at half, per Alibaba DashScope pricing as of 2026-10-18.
Top of the Fold
- DeepSeek updates its web, app and API default to V4.1-Flash-1017, selling faster runs at a new cost-performance frontier tuned for its own Harness and Tencent's CodeBuddy, company-stated per DeepSeek.
- Moonshot flags Kimi K3 as formally launched on its platform, pushing a new flagship tier into an already price-crushed field, per platform.moonshot.cn as of 2026-10-18.
- Otherwise a quiet morning up top.
Untranslated
- DeepSeek retires the legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names, silently rerouting those requests to V4.1-Flash-1017 at Flash pricing, a migration buried in its docs, per DeepSeek pricing notes as of 2026-10-18.
- 浙江大学's OmniAI team puts out EmbodiedMemo, a benchmark for whether embodied agents actually remember where an object was and that it later moved, preprint, as flagged by 新智元.
- Huawei is circulating plans for a new value-brand called 星芒 aimed at price-performance, circulating and unconfirmed, per 极客公园's morning roundup.
- Otherwise quiet across the Chinese community overnight.
The Backstory
- At Guangzhou's 140th Canton Fair, Guangzhou firm 动易科技 moved its humanoids from a 5-day badminton demo in April to a three-week run of bilingual tours plus table-tennis and heavy-transport booths, testing whether a trained skill survives a new task and a new floor, per 智东西.
The Kicker
- 虎嗅 reports Zhihu's homegrown 驽工理论, the layered-outsourcing theory that got banned on the platform, resurfacing as the plot logic of a 96%-fresh Resident Evil film, per 虎嗅.
