DEEPSEEK JUST SHIPPED V4.1-FLASH-1017 — BEATS OPUS 5 ON AGENTIC CODING AT $0.19 A TASK

DeepSeek has released DeepSeek-V4.1-Flash-1017, an update to its Flash model that the company says improves agentic and coding performance while cutting costs sharply. In DeepSeek's own benchmarks, the model scores 77.0 percent on DeepSWE at an average cost of just under $0.20 per task, and about 72 percent in its low setting for roughly $0.10. The predecessor scored 74.2 percent at around $0.50 per run.

DeepSeek says the gains come partly from reinforcement learning with self-feedback, which uses the model's own confidence scores and critiques as a reward signal for building complete software projects. The model is available through DeepSeek's website, app, and API, and is tuned for harnesses including DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Independent benchmarks had not yet appeared at the time of the report.