deepseek-v4.1VIDEO
AI daily roundup: V4.1-Flash claims DeepSWE 77 over Opus 5 but loses in Claude Code
A daily AI roundup says DeepSeek V4.1-Flash claims a DeepSWE score of 77 over Opus 5 at about $0.2 per task, but scores lower when run in Claude Code because it was trained for DeepSeek's own harness.
TellerAI-Leo
@leozhangaiArchived 11 days agoOriginally posted 2026.10.18
TellerAI-Leo @leozhangaiOriginally posted 2026.10.18
AI 早报 · 10.17｜Agent 开始组队 头条：DeepSeek V4.1-Flash 自报 DeepSWE 77 胜 Opus 5，每任务约 $0.2；但为自家 Harness 训，进 Claude Code 跑分反输。 另：Claude 半小时查清 NIST 曲线种子·GPT-6 Astra Pro 并行 Agent·4 个 Gemini Agent 27 小时通关 Minecraft #AI早报 #Agent
x.com/leozhangai/status/2111623417510191245Archived 2026.10.18 16:46