A New Report Says Moonshot and DeepSeek Trained on OpenAI's Model. It's Part of a Pattern.

The Information reported on October 29 that DeepSeek and Moonshot AI trained their newest models largely on output pulled from OpenAI's GPT-6 Astra, citing people with direct knowledge of the matter. The models named are Moonshot's Kimi K3.1 Pro, which shipped the week before, and DeepSeek's V4.1-Flash 1017, which launched October 17. Techzine reported that OpenAI, DeepSeek, and Moonshot did not immediately respond to requests for comment; The Next Web, in its own follow-up, said only that it had not independently verified the report. None of the outlets that have covered the claim say they've confirmed it themselves.

The mechanism being alleged is standard distillation: send a rival's model large volumes of prompts, collect its answers, and train a new model to reproduce them. Doing this at scale through a commercial API usually breaks that provider's terms of service rather than any copyright rule, which is why labs rarely pursue it in court and instead rely on account bans, rate limits, and public callouts. It's also, as OpenAI and Anthropic both argue, a way for a fast-following lab to approach a frontier model's behavior without paying for the compute that produced it.

A pattern, not a one-off accusation

This isn't the first time Moonshot has been in this position. OpenAI said on September 30 that it had linked a coordinated campaign, which began in July, to people associated with Moonshot AI. That campaign targeted the hidden reasoning traces behind model answers rather than the answers themselves, and OpenAI said it detected activity involving more than 15,000 accounts before shutting it down by July 28, banning the accounts, and sharing its findings with other labs. Earlier in 2026, OpenAI made a similar accusation against DeepSeek, saying affiliated accounts had been used to automatically harvest responses for training. Anthropic has described its own version of the same problem: it says it has disrupted distillation campaigns tied to seven China-based labs since February, that new safeguards cut the volume of data collected in these attacks by roughly 90 to 99 percent, and that it now actively protects reasoning transcripts and has permanently blocked tens of thousands of accounts.

Moonshot, for its part, has claimed, per The Next Web, that Kimi K3.1 Pro outperforms GPT-6 Astra on a coding benchmark. Nobody outside the two companies has verified that result either, which is the recurring problem with this entire dispute: the accusers and the accused are both grading their own work, and the labs with the most to gain from calling a rival's training process illegitimate are the ones its model is competing against for API revenue.

Why the timing stings

The report lands days after Kimi K3.1 Pro had a rough outing on Artificial Analysis's newly overhauled Intelligence Index v5, which knocked the model from fourth to ninth place overall even as it held on to the top spot among models with released weights. That benchmark reshuffle was its own story; what the distillation report adds is a specific, if unverified, explanation for how Kimi got competitive in the first place. A model that climbed the leaderboard partly by learning from a rival's outputs is a different kind of achievement than one that got there through its own pretraining and research, even if the benchmark score looks identical either way.

Weights for Kimi K3.1 Pro are due November 4, which means the open-weight community is about to get hands-on access to a model that two separate threads — an independent benchmark overhaul and an unverified distillation report — have each complicated in the same week. Neither development changes what the model can do. Both change how much credit it's reasonable to give Moonshot, specifically, for getting it there. The accusation is consistent with a pattern OpenAI and Anthropic have each separately documented since February, which doesn't make it true, but does make it the kind of claim that will keep resurfacing every time a Chinese lab's model lands close behind a frontier release, whether or not this specific instance is ever independently confirmed.