The Number
- DeepSeek's cheapest tier, deepseek-flash, now lists off-peak input at $0.003/MTok on a cache hit, half the peak $0.006 rate, per DeepSeek's pricing docs (2026-11-01).
- GLM-5.2 undercuts the field at $0.006/MTok prompt among China-lab models on OpenRouter, as of 2026-11-01.
- Kimi's K2 series stops taking new sign-ups from 2026-10-31, with full retirement planned for late December, per Moonshot's pricing page.
Top of the Fold
- DeepSeek's Harness team publishes a Super-agent architecture and an open research prototype, pitching cross-session, multi-task agents that work without being prompted, per its official post.
- Pika ships Pika H3 Flex, a MiniMax H3 derivative with open adapter weights and inference code for joint audio-video generation in six denoising steps, per 雷峰网.
- Ant Group's 灵光 2.0 repositions from a 30-second app builder to a managed personal assistant wired into Alipay, DingTalk, Feishu and Mijia, chasing Meta's Muse into Chinese consumer AI, per 极客公园.
Untranslated
- DeepSeek's Harness team also shipped an emergency patch to its Harness, blocking reverse-engineering of games, a quiet guardrail move drawing little English notice, per 量子位.
- 新智元 reports a Silicon Valley Chinese team, AFK AI (product Comma), shipped DeepSeek's Super-agent architecture back in April, months before the DeepSeek post, circulating and unverified, per 新智元.
- Westlake University stands up a World Model research center, three years after seeding the effort, per 量子位.
- Tsinghua's Yao Qizhi group publishes an algebraic world model paper aimed at letting AI reason more like Newton, a China-lab preprint signal, per 量子位.
The Backstory
- B站's comment sections filled with user-run AI accounts turning DeepSeek's blue-whale logo into a maid-dressed, rice-loving '大肥鱼' mascot, one video past 2 million views, with a counter-plugin built to filter the bots, per 极客公园.
The Kicker
- A veteran convenience-store vending robot, drilled over ten thousand times in the lab, still drops bottles at the shelf in the field, per 量子位.
