DeepSeek keeps competing on the price of finishing a job. On October 17 the Chinese lab released V4.1-Flash-1017, an open-weight update it calls a Pareto improvement, meaning better scores without higher cost.
The efficiency story is the headline. On the DeepSWE benchmark, the new model reaches 77 percent at just under $0.20 per task. Its direct predecessor hit 74.2 percent for about $0.50. DeepSeek’s own chart places GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 between $2 and more than $10 per task, without matching the smaller model’s accuracy.
The training method draws attention too. DeepSeek uses reinforcement learning with self-feedback, where the model treats its own confidence scores and self-written critiques as the reward signal, removing human raters and separate reward models from the loop.
Weights are on Hugging Face and ModelScope. The model is tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and the lab says it now builds whole projects, from Godot game recreations to 3D scenes.
Second half of the release: Search and Fetch APIs built on DeepSeek’s internal engine, which it says handles hundreds of millions of requests daily, priced at $0.75 per 1,000 calls with free access inside popular coding tools until October 31.
