DeepSeek V4.1 Flash 1017 发布：DeepSWE 77.0 超 Opus 5

Jurly @jurlycat原推发布于 2026.10.18

在 X 上打开原推

DeepSeek V4.1 Flash 1017 just landed on Oct 17. In just a few months, DeepSeek is already trading blows with frontier models on coding benchmarks. DeepSeek’s own DeepSWE v1.1 results: • Flash 1017: 77.0 • Claude Opus 5: 74.0 • GPT-5.6 Sol: 73.0 MIT weights. 1M context. Faster task completion and shorter outputs, according to DeepSeek. The official demos even include a Portal 2 recreation in Godot. Then there’s the API pricing per million tokens: • Off-peak: $0.15 input / $0.60 output • Peak: $0.30 input / $1.20 output Catching up this fast while keeping prices this low is wild. How do you compete with that? 😱