DeepSeek V4.1 Flash 官方跑分 11 项只赢 Opus 5 四项，真正优势是价格
梳理官方数据：agentic 四项（DeepSWE、Automation-Bench、Agents' Last Exam、APEX）胜出，HLE、Terminal-Bench、ProgramBench、NL2Repo 和视觉测试输给 Opus 5；MIT 开源权重，$0.30/$1.20 每百万 token，约为 Op
deepseek just dropped v4.1 flash 1017, and the headline everyone's sharing is "it beats claude opus 5." that's half true. in deepseek's own tests it wins 4 of 11 benchmarks against opus 5, mostly agentic ones: deepswe, automation-bench, agents' last exam and apex bench. it loses the other 7. hle, terminal-bench, programbench, nl2repo and the vision tests all go to opus. and opus 5 isn't even anthropic's top model anymore. opus 5.5 and fable 5.1 are already out. so the "flash beats the flagship" story doesn't hold. the real story is price. this is an open-weights model under mit. it costs $0.30 in and $1.20 out per million tokens. opus 5 costs $5 and $25. that's roughly 17-20x cheaper. on deepseek's own chart it solves a deepswe task for under $0.20. rivals land somewhere between $2 and $10+ per task. one catch: every number here is self-reported. independent tests of this version aren't out yet. if they hold up, the question stops being "is it smarter than claude." it becomes "why pay 20x more for the tasks where it's close enough."