DeepSeek V4.1 Flash 1017: Terminal-Bench 31.2 to 42.3, but Opus 5 still leads 7 of 11 rows

PixelShipper @PixelShipperOriginally posted 2026.10.18

Open original on X

DeepSeek called it a minor update. Terminal-Bench went from 31.2 to 42.3. V4.1 Flash 1017 dropped this morning: • Up on all 11 rows of DeepSeek's own table • ProgramBench 20.3 → 29.1, Apex Bench 64.6 → 72.1 • Same API price: $0.15 in / $0.60 out per 1M tokens • MIT license, 1M context, weights on Hugging Face The catch: on that same table, Claude Opus 5 still leads 7 of 11 rows. And Opus 5 isn't Anthropic's current model.