TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
The Information, as reported by The Decoder, says DeepSeek and Moonshot AI relied on outputs from OpenAI’s GPT-6 Astra to train newer models after Anthropic strengthened defenses against attempts to extract model data. The labs’ alleged use of Astra, and the details of any training, have not been independently confirmed in the supplied reporting; OpenAI and Anthropic prohibit distillation that exposes hidden reasoning.
DeepSeek and Moonshot AI have relied heavily on outputs from OpenAI’s GPT-6 Astra to train newer models after Anthropic tightened defenses around Claude, according to The Information, as reported by The Decoder. The alleged shift focuses attention on how frontier-model companies can limit competitors’ use of their systems for distillation, a practice that may involve extracting a model’s reasoning or other behavior through its outputs.
People familiar with the matter told The Information that the models trained using Astra outputs include Moonshot’s Kimi K3.1 Pro and versions identified as V4.1-Flash 1017. The report describes the labs as leaning heavily on Astra, but the supplied account does not quantify how many outputs were collected, how they were used, or how much they contributed to the models’ training.
OpenAI and Anthropic both prohibit distillation that involves using their models to reveal hidden reasoning steps. The reporting says Claude had been a main target before Anthropic strengthened its defenses. Anthropic says its newer methods cut the amount of data attackers can extract by 90% to 99%; that figure is the company’s own claim, and the source material provides no independent assessment or measurement details.
Moonshot declined to comment to The Information, while DeepSeek did not respond to a request for comment, according to The Decoder’s account. OpenAI previously said it had identified a coordinated distillation campaign in late September that it said involved people linked to Moonshot. That allegation does not, on its own, confirm the scope or methods of the activity described in the newer report.
Why Astra Is a Distillation Target
The reported movement from Claude toward Astra suggests that stronger extraction defenses can change which frontier models are targeted without ending attempts to obtain useful training data. If labs can collect a competing system’s outputs at scale, they may use those responses to shape a model’s behavior without building the same capabilities entirely from scratch. The practical effect depends on the volume, quality and role of the data, none of which is specified in the available account.
For model developers, the issue reaches beyond policy enforcement. Reasoning traces and generated answers can reveal information that makes a system’s behavior easier to imitate, while tighter restrictions may reduce the usefulness of an API for legitimate as well as unauthorized research. For users and customers, the dispute could affect access rules, monitoring and the reliability of claims about how a model was trained.
The report also highlights a challenge for companies trying to protect valuable models: defenses may limit data extraction from one system while activity shifts to another. That is a reported pattern, not proof that every attempt succeeds or that Astra’s outputs materially determined the named models.
Products Worth Considering
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Claude to Astra
The Decoder’s account says Claude had previously been the go-to model for this kind of activity, but that Anthropic has since tightened its defenses. Anthropic describes those measures as reducing the data an attacker can obtain by 90% to 99%. The supplied material does not describe the technical methods behind the safeguards or identify the testing conditions for the claimed reduction.
In late September, OpenAI reported a coordinated distillation campaign that it said involved people linked to Moonshot. The company said attackers adjusted their tactics whenever it introduced a new defense. This was OpenAI’s characterization of the campaign; the source material does not include independent verification, a response from Moonshot about that specific allegation, or a detailed account of the evidence.
The current report adds a claim about Chinese labs’ use of Astra outputs and names models said to have been trained with them. The timing and relationship between those activities and the campaign OpenAI described in September are not established in the supplied reporting.
“Its new methods cut the amount of data attackers can pull by 90% to 99%.”
— Anthropic
Evidence Behind the Training Claims
The reported use of Astra outputs rests on information from people familiar with the matter, according to The Information; the supplied account does not provide public documentation, data volumes or technical evidence showing how the outputs were incorporated. DeepSeek did not respond to a request for comment, and Moonshot declined to comment.
It is also unclear what “leaned heavily” means in measurable terms, whether the named model versions used hidden reasoning traces specifically, and how much any such material affected performance. The source says distillation of this kind requires models to reveal hidden reasoning, but it does not establish the precise prompts, access methods or datasets involved in the cases it describes.
OpenAI’s September account and Anthropic’s stated defense improvements are company claims. The material does not establish whether the activity reported by The Information violated specific terms in a way that has been adjudicated, or whether any legal or contractual action has followed.
Responses and Further Evidence
The next meaningful developments would be responses from DeepSeek and Moonshot, or further reporting that documents the scale and methods of the alleged Astra use. OpenAI may also provide details about whether it has identified activity tied to the models named in the report or changed access controls in response.
Anthropic’s claimed reduction in extractable data would be easier to evaluate if the company or an independent source published the testing method, baseline and conditions. Until then, the reported 90% to 99% reduction should be treated as Anthropic’s description of its safeguards, not as a verified result across all extraction attempts.
For now, the account points to a shifting contest over model outputs and defensive measures. It does not establish that the named Chinese models were built primarily from Astra, how the companies will respond, or whether tightened defenses will stop similar efforts from moving to other systems.
Key Questions
What does the report say DeepSeek and Moonshot did?
According to The Information, as reported by The Decoder, both labs relied heavily on OpenAI’s GPT-6 Astra outputs to train newer models. The account names Kimi K3.1 Pro and versions identified as V4.1-Flash 1017, but does not quantify Astra’s contribution.
What is model distillation?
Distillation is a way to use one model’s outputs to help train another model. In the activity described by the report, the concern is that outputs could include hidden reasoning steps that make the original system’s behavior easier to imitate.
Why did attention shift from Claude to Astra?
The report says Claude had been a common target, but Anthropic tightened its defenses. Anthropic says those methods reduced the data attackers could extract by 90% to 99%; the account does not independently verify that figure.
Have DeepSeek or Moonshot confirmed the allegations?
No confirmation is included in the supplied reporting. Moonshot declined to comment and DeepSeek did not respond to a request for comment, according to The Decoder.
What is still unknown about Astra’s use?
The reporting does not specify the number of outputs collected, the exact methods used, or how much those outputs influenced the named models. It also does not establish whether the activity described is connected to OpenAI’s separate late-September account of a campaign it said involved people linked to Moonshot.
Source: the-decoder.com
