DeepSeek and Moonshot Reportedly Used OpenAI's GPT-6 Astra to Train Their Latest Models
The Information says Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017 benefited from distilling Astra after Anthropic tightened its defenses on Claude.
Published
Last updated / Fact-checked on
According to an October 28, 2026 report from The Information, Chinese AI developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra to train their latest models. The report names Moonshot's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017. Neither company has confirmed the claims.
Image: Farzad Nazifi (Unsplash) / CC0 1.0, via Wikimedia Commons
What the report says
The Information cited people with direct knowledge of the matter. According to summaries of the paywalled article, it attributed the Moonshot claim to three people and the DeepSeek claim to two.
| Model | Developer | Release | Reported use of Astra |
|---|---|---|---|
| Kimi K3.1 Pro | Moonshot AI | The week before the report | Distilled Astra outputs during development |
| V4.1-Flash 1017 | DeepSeek | Oct 17, 2026 | Drew on Astra for the upgraded model |
OpenAI released GPT-6 Astra on September 3, 2026. The company calls it its most capable model. The report says Chinese labs had focused on Anthropic's Claude until recently, then shifted to Astra after Anthropic introduced stronger countermeasures.
The Decoder, summarizing the report, said Moonshot declined to comment and DeepSeek did not respond. Reuters said OpenAI, DeepSeek and Moonshot did not immediately respond to its own requests for comment.
What distillation is
Distillation trains one model on the outputs of a stronger one. It can transfer capabilities more quickly and cheaply than training from scratch. OpenAI and Anthropic use it to build their own smaller models. Both companies forbid outside firms from distilling their models.
According to The Decoder, distillation of this kind works only if a lab can get the target model to reveal its hidden reasoning steps. That is something both OpenAI and Anthropic prohibit.
Reported use of a model's outputs is not evidence that model weights or internal training data were taken. How much Astra actually contributed to either Chinese model has not been independently established.
Anthropic's defenses and OpenAI's earlier findings
Anthropic said earlier in October that new measures had cut the data harvested in distillation attacks by about 90% to 99%. According to The Decoder's October 13 report:
- the company now restricts access to Claude's internal reasoning
- it runs a classifier to detect extraction attempts
- it automatically bans flagged accounts, and tens of thousands have already been banned
Anthropic acknowledged that some legitimate users have been caught up in the process. These figures are Anthropic's own and have not been independently audited.
OpenAI has also been tracking campaigns. In a September 30 blog post, it said it had linked the core of a distillation campaign that began in July to individuals associated with Moonshot AI. The Decoder reported that an OpenAI spokesperson said the company takes the problem very seriously, but attackers have adjusted their tactics each time it rolled out a new defense.
Why it matters
Distillation has become a flashpoint as US-China competition in AI intensifies. Anthropic, OpenAI and Google argue that the practice lets rivals reproduce advanced capabilities faster and at lower cost.
The reported move from Claude to Astra suggests that a defense at one lab can push extraction toward another lab rather than end it. For frontier developers, that raises the stakes on industry-wide detection and on how much hidden reasoning their APIs expose.
For companies evaluating open-weight models, the report is a reminder that a model's performance does not show where its capabilities came from. That provenance is hard to verify from the outside.
Until DeepSeek, Moonshot or OpenAI address the claims directly, they remain allegations attributed to unnamed sources. More on AI competition is in our Artificial Intelligence hub.
How we reported this
The Information's article is paywalled, so we relied on Reuters' summary and The Decoder's coverage and attribute the claims to the report. We confirmed Astra's release date with OpenAI. We cut a paraphrased OpenAI statement that we couldn't confirm word for word, along with favicon clutter and tracking links.
Sources
- The Information: OpenAI's Astra Model Becomes Top Distillation Target for China's AI Labs, Oct 28, 2026 (paywalled).
- Reuters via Investing.com: DeepSeek, Moonshot used OpenAI's GPT-6 Astra to train latest models, The Information reports, Oct 28, 2026.
- The Decoder: OpenAI's Astra becomes the main target for Chinese AI labs after Anthropic tightens its defenses
- The Decoder: Anthropic says it cut Claude reasoning token theft by up to 99%, Oct 13, 2026.
- OpenAI: GPT-6 Astra, Sep 3, 2026.
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.