V4.1 Flash 官方跑分：修 bug 胜 Opus 5，终端任务落后
DeepSWE 77.0 对 74.0、Automation-Bench 60.5 对 50.3 领先，Terminal-Bench 42.3 对 51.8、NL2Repo 68.2 对 75.3 落后；DeepSWE 单任务约 $0.19，Opus 5 约 $12。
holy sh*t, deepseek just shipped opus 5 level bug fixing for $0.19 a task v4.1 flash-1017, open weights, in deepseek's own table vs opus 5: → deepswe: 77.0 vs 74.0 → automation-bench: 60.5 vs 50.3 → terminal-bench: 42.3 vs 51.8 → nl2repo: 68.2 vs 75.3 it wins on bug fixes and loses on terminal work their cost chart puts a deepswe task at ~$0.19 on flash-1017 vs ~$12 on opus 5 at max effort the price, per 1M tokens: > $0.15 in / $0.60 out off-peak > $0.30 / $1.20 at peak (01-04 and 06-10 utc on weekdays) > 1M context, 384K output, anthropic-format api so don't migrate. add a lane: → bug fixes and repo patches: deepseek-flash → terminal work and plans: keep opus → search and labels: haiku 5.5 → anything irreversible: you these are deepseek's numbers on harnesses deepseek picked. run 20 of your own tasks through both before you move a single route
这条是引用着另一条写的：x.com/deepseek_ai/status/2111375640717033777