The Number
- Qwen3.8-Max lists at 12 RMB per million input tokens and 36 RMB output on DashScope, unchanged across its 0902 and 3.7 variants, per Alibaba's pricing page (2026-10-25).
- DeepSeek folds deepseek-v4-flash into DeepSeek-V4.1-Flash-1017, retiring the old Flash names and billing all at the Flash price, per DeepSeek docs (2026-10-25).
- Kimi K2 series stops onboarding new users from Oct 31, 2026 and goes fully offline by end-December, per Moonshot's pricing docs (2026-10-25).
Top of the Fold
- Otherwise a quiet morning up top; no marquee China-AI story cleared the wires overnight.
Untranslated
- Moonshot sets a hard sunset on its K2 line, blocking new signups from Oct 31 and killing all K2 models by December as it pushes K3.1 Pro, per Moonshot pricing docs.
- Alibaba's US-region Qwen3.8-Max carries a premium, 14.988 RMB per million input against 12 RMB globally, a split rarely noted in English coverage, per the DashScope price table.
- Tsinghua, Oxford and Stanford researchers document "Thinking Inertia," models that keep reasoning before an answer even with thinking mode off, worse on open-ended prompts, per a NeurIPS'26 preprint flagged by 新智元 (company-stated, unverified).
The Backstory
- DeepSeek quietly opens its agent stack, putting DeepSeek Harness into developer preview for harness builders worldwide, per DeepSeek docs.
The Kicker
- Chinese broadcasters are on an AI-compute buying spree, with 259 stations and media units tendering 92.28M RMB in recent deals and the standout line item simply labeled "tokens," per 娱乐资本论 via 虎嗅.
