DeepSeek and Moonshot trained on OpenAI Astra outputs, report says — after Claude got harder to copy

The target of Chinese AI copying has moved from Claude to GPT, according to a new report. The Information reported on October 28, 2026 (US time), citing people with direct knowledge, that DeepSeek and Moonshot AI leaned heavily on outputs from OpenAI's GPT-6 Astra when building the models they released in October. Reuters carried the report.

Two models are named: DeepSeek's V4.1-Flash 1017, released October 17, and Moonshot's Kimi K3.1 Pro, released October 23. Both were widely described as close to the best US models. If the report is right, part of that catch-up came from training on Astra's answers. No evidence has been published, and neither Chinese company has confirmed it.

1. What was reported

This practice, often called a distillation attack, uses a rival model's answers as training data. The underlying technique, model distillation, is routine: labs use it to shrink their own large models into cheaper ones. The dispute is over doing it to someone else's model without permission.

2. Why Astra — how the target shifted

The sources' account: Chinese labs had mostly targeted Anthropic's Claude. When Anthropic introduced strong countermeasures, they moved to OpenAI, which they saw as easier to work around.

Anthropic's disclosed results are large:

When one door is reinforced, attackers try the next one. That shift is the core claim of the report. Neither OpenAI nor Anthropic offers its models in China, and both forbid other companies from distilling them. OpenAI's terms generally bar customers from using outputs to develop competing AI models.

3. The two models named

AI Weekly reported that V4.1-Flash 1017 rose from 70.9 to 75.0 on a third-party browser-operation test, close to Astra's 77.9, at about $0.22 per task against Astra's $10.77 — roughly 49 times cheaper. Only one outlet carried these numbers. The pattern they describe — near-frontier scores at a small fraction of the price — is why distillation disputes are bound up with the AI price war.

4. What OpenAI already disclosed in September

This is not the first such claim. On September 30, OpenAI said it had disrupted a coordinated reasoning-extraction campaign that began in July.

OpenAI warned that extracting protected reasoning could let competitors reproduce advanced problem-solving while skipping safety filters. That episode happened before Astra's September 3 launch. The new report claims the same activity continued against Astra.

5. Why it matters

For readers in Korea and elsewhere outside the US, Chinese open-weight models have become attractive because they are close in quality and far cheaper. The allegation adds a new question to that choice: whether the model's training data was obtained in breach of another provider's terms.

Distillation itself is not illegal; OpenAI and Anthropic both use it internally. The questions are permission, scale and terms of service — and those are hard to prove from outside. Similar answers from two models do not by themselves show that one was trained on the other.

6. What remains unknown

• Evidence: The report relies on anonymous sources. No logs, account data or training sets are public.
• Company positions: DeepSeek and Moonshot have not confirmed. Moonshot has previously denied distillation.
• OpenAI's next step: Whether OpenAI will publish account counts or attribution details, as it did in September, is unknown.
• Performance and cost numbers: The browser-test scores and per-task costs come from a single outlet.
• November 4: Whether Moonshot releases the Kimi K3.1 Pro weights on schedule is the next checkpoint.