DeepSeek, Moonshot Trained New Models on OpenAI's GPT-6 Astra
TL;DR
- DeepSeek V4.1-Flash 1017 and Moonshot's Kimi K3.1 Pro were trained on outputs from OpenAI's GPT-6 Astra, The Information reports.
- DeepSeek's browser-operation score moved from 70.9 to 75.0, near Astra's 77.9, at $0.22 per task versus Astra's $10.77.
- OpenAI's September 30 post said it had linked a core distillation campaign begun in July to individuals associated with Moonshot AI.
Chinese AI developers DeepSeek and Moonshot AI "relied heavily on outputs from OpenAI's powerful GPT-6 Astra model to train their latest models," The Information reported. Moonshot's Kimi K3.1 Pro shipped last week. DeepSeek's upgraded V4.1-Flash 1017 followed on October 17.
Distillation is the technique: query a stronger model at scale, collect its answers, train a cheaper one on them. In DeepSeek's case the payoff shows up in benchmarks. On a third-party browser-operation test, V4.1-Flash 1017 moved from 70.9 to 75.0, within reach of Astra's 77.9. The per-task cost on Astra: $10.77. On the DeepSeek update: $0.22. A nearly 49-fold difference.
OpenAI has started naming names. In a September 30 blog post, the company said it had "linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI." Anthropic said this month that new defenses had cut the data harvested through such attacks by roughly 90 to 99 percent, citing Zhipu's GLM-5.3 and tying a separate campaign to Alibaba's Qwen.
OpenAI, DeepSeek and Moonshot AI did not immediately respond to Reuters requests for comment.
Originally reported by theinformation.com
Read the original article →Original headline: The Information: DeepSeek V4.1-Flash and Moonshot's Kimi K3.1 Pro Were Trained on GPT-6 Astra Distillation