DeepSeek V4.1 Flash 1017: Cheaper, Faster Coding Agent With New Search API

On October 17, DeepSeek released DeepSeek V4.1-Flash-1017, a minor version update that keeps the same base model calling convention and pricing structure. The model weights are open-sourced on Hugging Face ( https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope (

https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017

).

Benchmark Gains and Pareto Frontier

Official benchmarks show significant improvements across reasoning, coding, agent tasks, and visual benchmarks. The model shortens task execution time and output length, letting users complete tasks faster and at lower cost. DeepSeek claims the new version advances the performance-cost Pareto frontier in multiple agent evaluation frameworks (Harness environments), meaning performance improves without sacrificing cost efficiency.

Key benchmark results (DeepSeek's self-reported figures, using the mini-swe-v2 framework):

DeepSWE v1.1 (real-world software engineering benchmark: given a real repo and issue, measure if the generated patch solves it end-to-end): 77% pass rate at the max tier, surpassing GPT-6.1 Sol, GPT-6 Astra, and Claud Fable 5 — larger flagship models.

Automation-Bench v1.0.6 , Agents' Last Exam , Apex Bench v1.1 : all exceed Claude Opus 5 and other listed models.

Chartography (chart understanding with tools): score approaches Claude Opus 5, exceeds other listed models.