DeepSeek V4.1 Flash 1017: better agent loops, Search at ¥5 per 1K calls
The author notes 1017 wanders less in agent loops, is already routed behind deepseek-flash, was optimized for DeepSeek Harness and Tencent CodeBuddy/WorkBuddy, and prices Search at ¥5 per 1,000 calls, while still splitting architect/challenger/reviewer roles to curb hallucination.
a few more things i’ve noticed with DeepSeek V4.1-Flash-1017: 1. the new checkpoint seems much better suited to agent loops. i’m seeing fewer cases where it wanders before taking the next concrete action. 2. the updated model is already routed behind deepseek-flash, so i didn’t need to change my existing API config to start using 1017. 3. DeepSeek specifically optimized this release for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, which makes the harness choice matter more than before. 4. the new Search API costs CNY 5 per 1,000 calls, or around $0.70, which is cheap enough that i can use search much more aggressively inside coding agents. 5. one thing i’m watching closely is hallucination rate. 1017 feels more capable, but i still get better results when i separate the workflow into architect → challenger → reviewer instead of letting one agent approve its own work. DeepSeek is improving the full agent stack around it.
This post quotes another post: “DeepSeek V4.1-Flash-1017: first hands-on observations”