DeepSeek and Moonshot are said to have trained on OpenAI's top model.
An unconfirmed report says two Chinese labs trained their newest models on answers from OpenAI's GPT-6 Astra. The Information, citing unnamed sources, says they switched to OpenAI after Anthropic made copying Claude much harder. Moonshot declined to comment and DeepSeek did not respond, so the claim remains unverified.
Why it mattersIf true, bans on distillation, meaning training on a rival's answers, are not stopping it. Cheap Chinese models may owe part of their skill to the US models they compete with.
⚠ Unverified rumor — treat with caution
Read in the app — free, in 9 languagesRelated stories
Screenshots suggest Kimi asked influencers to hide paid posts.2026-10-25Developers think Anthropic's next top model is only days away.
2026-10-21Anthropic says it cut the copying of Claude's reasoning by 90%.
2026-10-13Google is said to be testing a Gemini 4 version that rivals Claude.
2026-10-10Developers think a free mystery model belongs to Thinking Machines.
2026-10-06