Why Did DeepSeek's New Version Suddenly Get Stronger? Foreign Media Expose Astra Distillation Inside Story, Liang Wenfeng Compromises Under Competitive Pressure

Beating AI News Flash: DeepSeek's new model capabilities have surged, and Astra may truly deserve the credit behind it.

The Information, citing multiple people with direct knowledge, reported that DeepSeek's V4.1-Flash-1017 updated on October 17, as well as Moonshot's newly released Kimi K3.1 Pro, both extensively used outputs from OpenAI GPT-6 Astra for distillation during training.

The new DeepSeek version performs especially well on browser operation tasks. In Browser Use's 180 real-world web task tests, its score rose from 70.9 in the old version to 75.0, approaching Astra's 77.9. But the model invocation cost per question is only $0.22, while Astra's is as high as $10.77, a difference of nearly 49 times.

Previously, Claude had always been the main distillation target for domestic AI companies. But Anthropic has been tightening protections over the past two months, making it significantly harder to extract the model's reasoning process, so domestic labs turned their attention to Astra.

DeepSeek founder Liang Wenfeng was originally unwilling to rely on distillation in V4 development. But after seeing domestic competitors such as Zhipu GLM make progress through distillation, he ultimately yielded to competitive pressure.