Report Says DeepSeek and Moonshot Used GPT-6 Astra Outputs to Train New Models

DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra when training their latest models, according to a report citing people familiar with the matter. The claim adds a new dimension to disputes over model distillation, as Moonshot also reports a coding-agent benchmark result above Astra’s. Reuters, relaying The Information’s report, identified Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017 as models that benefited from training on Astra’s outputs. Reuters said it could not immediately obtain comments from OpenAI, DeepSeek or Moonshot AI.