DeepSeek V4.1-Flash-1017 beats larger competitors on agent tasks, and per-task cost drops sharply again

DeepSeek released DeepSeek-V4.1-Flash-1017, claiming it surpasses larger models such as Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3 on coding and agent tasks. The model was post-trained with reinforcement learning with self-feedback (RLSF) and has open-sourced weights; it is available on the official website, app, and API. In self-tests on DeepSWE it reaches up to 77.0% at less than $0.20 per task, and 72% at about $0.10 in the low-configuration setting, while the previous generation was 74.2% at about $0.50. DeepSeek also opened its internal search engine's Search & Fetch API to developers for the first time, at $0.75 per 1,000 calls.