The cheapest way to catch up with a rival, it seems, is to ask it a great many questions and keep the answers. A media report citing people familiar with the matter says DeepSeek and Moonshot AI did exactly that with GPT-6 Astra, using its outputs to train their newest models.
The models named are Kimi K3.1 Pro, which Moonshot released last week, and an updated DeepSeek V4.1-Flash. Everything here rests on anonymous sources, so treat it as an allegation and not a finding.
What the report claims
According to the sources, both companies used distillation. The idea is simple. A developer sends a large volume of prompts to a stronger model, collects the replies, and then trains a different system on that material.
The appeal is arithmetic. You reproduce part of what the stronger model can do without building those abilities from scratch, and without paying anything like what the original training cost.
The material available here includes no comment from DeepSeek or Moonshot.
Not the first time Moonshot has come up
OpenAI has already accused Moonshot of something similar. On September 30, the company said users linked to the developer of Kimi had run a campaign in July to pull hidden reasoning out of its models' responses.
The numbers OpenAI gave:
- Scale: it found related activity from more than 15,000 users.
- Response: by July 28 it had blocked the accounts involved and stopped the campaign.
- Follow-up: it passed what it had collected to other AI model developers.
The new report does not say whether the two episodes are connected. It simply lands on top of the earlier one.
Why the big labs care
OpenAI, Anthropic and Google all hold the view that distillation hands rivals access to frontier capabilities without the matching investment. American agencies have also pointed at Chinese companies they believe tried to extract data from US AI models.
So the dispute is about competition between American and Chinese developers, and about who pays for progress and who gets to borrow it.
Defenses are being built
The labs are not sitting still. In early October, Anthropic said new protective measures had cut the data collected from Claude models in distillation attacks by roughly 90 to 99 percent.
That is a large range, and Anthropic's own figure, so it is worth reading as the company's claim about its own defenses. It does show that blocking this kind of harvesting is now treated as an engineering project with a metric attached.
The performance claims remain unchecked
Both Chinese models scored well in their developers' internal tests. Moonshot says Kimi K3.1 Pro beat GPT-6 Astra on one coding benchmark. DeepSeek positioned V4.1-Flash as a cheaper alternative to American models.
None of these results has been independently confirmed. A model that beats its teacher on one test is a pleasing story, though a single self-reported benchmark does not settle how the two compare in general.
A date that does not line up
The source text is inconsistent about DeepSeek's timeline. It gives October 17 as the date the updated V4.1-Flash was presented. A few paragraphs later it says DeepSeek introduced V4.1-Flash in September as a cheaper option.
The two statements can both be true if September was the original launch and October the update. The report does not say that, so the exact timing stays unclear.
The awkward part
If the account is accurate, it exposes a problem every leading lab shares. A model has to answer the public's questions to be a product at all. Anyone who can phrase those questions at scale gets the answers too, and a determined competitor can turn them into training data.
The labs have a business built on being asked things, and they would now like to be asked slightly fewer of them by the wrong people.