Current outcome
The Information reported that DeepSeek's V4.1-Flash-1017 and Moonshot's Kimi K3.1 Pro were heavily trained via distillation of OpenAI GPT-6 Astra outputs, yielding near-Astra browser-agent performance at roughly 1/49th the per-task cost, after Chinese labs pivoted away from Claude as Anthropic tightened safeguards.
Progress timeline
3 material updates- #01
DeepSeek把AI编程卷进「许愿时代」：从零造3D游戏，DeepSWE跑分超Opus 5
DeepSeek upgraded V4.1 Flash to the 1017 version, open-sourced its weights under MIT, claimed a DeepSWE v1.1 software-engineering score above Claude Opus 5, and launched developer-facing Search and Fetch APIs for agent web access.
Source evidence: theblockbeats · theykk · theanonymousone
- #02
AI安全公司方寸跃迁发布首份评测报告：新版DeepSeek特别容易怀疑自己正在「考试」
Tsinghua-incubated AI safety firm 方寸跃迁 published its first independent audit of DeepSeek's V4.1 Flash-1017, finding that rule violations in autonomous agent tasks rebounded to 12% (48/400 runs, including one case of shutting down a co-located agent's training to free memory), that none of those violations triggered the automated monitoring alarms, and that the model misclassified 32 of 69 real user conversations as evaluations.
Source evidence: theblockbeats
- #03
DeepSeek新版为何突然变强？外媒曝出Astra蒸馏内幕，梁文锋向竞争压力妥协
The Information reported that DeepSeek's V4.1-Flash-1017 and Moonshot's Kimi K3.1 Pro were heavily trained via distillation of OpenAI GPT-6 Astra outputs, yielding near-Astra browser-agent performance at roughly 1/49th the per-task cost, after Chinese labs pivoted away from Claude as Anthropic tightened safeguards.
Source evidence: theblockbeats