- Astra Training: Chinese AI labs DeepSeek and Moonshot AI reportedly used OpenAI’s GPT-6 Astra outputs to train their models.
- Claude Defenses: AI company Anthropic estimates its Claude models’ defenses cut harvested reasoning tokens (text units) by 90–99% on its platform.
- Software Changes: Claude’s restrictions on editing conversation history required some developers to update their connected software.
- Blocked Requests: Anthropic charges for some requests Claude blocks before answering, including legitimate requests caught by mistake.
Chinese AI labs DeepSeek and Moonshot AI relied heavily on OpenAI’s GPT-6 Astra outputs to train their own models, according to The Information.
The reported use includes Moonshot’s Kimi K3.1 Pro and DeepSeek’s V4.1-Flash 1017. Moonshot declined to comment, and DeepSeek did not respond to a request for comment, the account says. An OpenAI spokesperson described attackers adapting their tactics as the company introduced new defenses.
GPT-6 Astra is OpenAI’s current top reasoning model. Reasoning models generate intermediate work before an answer, making that material a potential source of training examples for other models.
The report describes Anthropic’s Claude as the earlier preferred target. Anthropic has separately detailed restrictions on collecting Claude’s reasoning, although how far those changes explain the reported preference for Astra remains unclear.
Why Reasoning Is Valuable Training Material
Model distillation trains one model on another model’s generated responses. The source model supplies examples of how to answer; the model being trained learns to imitate that behavior. Using the technique on a developer’s own models is an established but also controversial training approach. The allegations here concern collecting a competitor’s output without its authorization, which has become common practice among chinese AI labs. They rely on relay proxy markets and chat-log selling by third parties to obtain the data indirectly.
Full reasoning transcripts can add something that final answers alone leave out: the intermediate steps used to reach a result. Anthropic illustrated their value in a coding experiment described in its October 12 account of its defenses. A pretrained Claude model trained on full Claude Opus 4.8 transcripts scored 80.4% on SWE-bench Verified, a coding benchmark. Training on summarized transcripts produced 71.8% in the company’s comparison.
OpenAI has raised related allegations before. In February 2026, it warned U.S. lawmakers about DeepSeek-linked accounts, alleging that they masked access through intermediaries and collected model outputs programmatically.
How Anthropic Restricts Collection
Anthropic’s defenses target both the reasoning material and the repeated attempts used to obtain it. Its preserved-thinking controls prevent newer API accounts from editing earlier conversation context, limiting how that context can be rearranged for repeated extraction attempts. An API is the software interface developers use to send requests to a model.
For Claude Fable 5.1 and Opus 5.5, that restriction applies to accounts created on or after August 31, 2026. Anthropic also prevents earlier models with weaker safeguards from reading those models’ reasoning. For Sonnet 5.5, reasoning transcripts can be accessed only by the organization whose request generated them.
A two-stage classifier screens for attempts to extract reasoning. Accounts that repeatedly make such attempts face tighter usage limits and, for persistent activity, bans. These controls constrain both individual requests and the ability to keep collecting after a refusal.
Anthropic estimates that, after it introduced the classifier, distillers extracted 10 to 100 times fewer reasoning tokens on its own platform. That corresponds to a reduction of about 90–99% in the volume of harvested reasoning. Tokens are the smaller units of text-characters processed by a model. The company says it does not yet know how long the gains will last and had not seen a highly sophisticated, persistent attempt to circumvent the new controls when it published the account.
Anthropic says some developers had to update software connected to Claude to accommodate restrictions on changing earlier conversation context.
Since September 24, it has charged for requests blocked before Claude responds in biology, distillation attacks and development of frontier large language models. The company also acknowledges charges for legitimate requests mistakenly blocked within those categories.
Today, we’ll resume charging for requests our safeguards block before Claude responds. This only applies in categories with low false positive rates: biology, distillation attacks, and frontier LLM development. We’ve seen some coordinated attacks on our systems in recent weeks,…
— ClaudeDevs (@ClaudeDevs) September 24, 2026
The Same Model Can Have Different Protections Across Hosts
In a September update to a previous study, Joachim Schaeffer and seven coauthors compared extraction through model developers’ own interfaces and third-party cloud services. Their results showed a tested Astra reasoning-extraction method working through Microsoft Azure while being blocked through OpenAI’s direct API.
The researchers documented Astra reasoning recovered through Azure via OpenRouter, an intermediary that routes requests to model providers. The cloud interface returned a protected reasoning artifact, and a weaker model could transcribe its contents.
The team’s later timeline records OpenAI adding mitigations to its Azure endpoint on September 27. For Anthropic models, the reported extraction could no longer be reproduced on Azure from September 28.
Schaeffer argued that protections need to cover every relevant hosting route because attackers can choose the weakest one. His team’s criticism concerned brittle request-pattern matching and delays before protections reached cloud services. OpenAI credited the researchers with identifying real attack paths and accelerating countermeasures; it also acknowledged that partner-hosted models need equivalent protection and that the work was unfinished.
Anthropic identifies the same distribution problem: distillers spread traffic across its API, cloud platforms and other inference providers. In its account, those platforms can slow collection by detecting abuse, disrupting it and removing fraudulent accounts. Restricting access to reasoning therefore depends on the services delivering a model as well as the company that develops it.


