Brief, mostly signal

DeepSeek V4.1-Flash-1017 cuts per-task cost on agent benchmarks

Vendor-reported 77.0% on DeepSWE at under $0.20 per task, versus $0.50 for the prior version. Weights are open on Hugging Face and ModelScope.

Read the source at egamers.io

What the research found

DeepSeek V4.1-Flash-1017 scores 77.0 on DeepSWE versus 74.2 for the prior version at under $0.20 per task versus about $0.50; these are vendor-only figures, tuned for DeepSeek Harness v0.2.1-rc.1 or later, and weights are open on Hugging Face and ModelScope.

Claims and how well the evidence supports them

DeepSeek V4.1-Flash-1017 scores 77.0 on DeepSWE versus 74.2 for the prior version 98%
Cost per task is under $0.20 versus about $0.50 for the prior version 97%
Results are vendor-reported and tuned for DeepSeek Harness v0.2.1-rc.1 or later 96%

Evidence

egamers.io https://egamers.io/deepseeks-v4-1-flash-1017-outscores-larger-models-on-agentic-benchmarks-while-dropping-per-task-costs-once-more
note.com https://note.com/kawara_works/n/nf183b1e4833b?hl=en

Actionability rated medium.