DeepSeek and Moonshot AI leaned heavily on OpenAI's GPT-6 Astra to train their newest AI models, The Information reported Wednesday, citing people with direct knowledge of the matter. The outlet named Moonshot's Kimi K3.1 Pro, released last week, and an updated version of DeepSeek's V4.1-Flash, which launched on October 17, as the products in question. The Next Web, which first surfaced the report, said it has not independently verified the claims.
The technique at the center of the report is called distillation. A company sends huge volumes of prompts to a rival's model, collects the answers, and uses that output to train its own system. Done at scale, it lets a smaller or newer model inherit chunks of a more advanced one's behavior without the enormous cost of building those capabilities from the ground up.
This isn't the first time Moonshot has faced this accusation. On September 30, OpenAI said it had caught people linked to the Kimi developer running a coordinated effort in July to pull the hidden reasoning steps behind its models' answers, rather than just the final output. OpenAI said the activity touched more than 15,000 user accounts before it shut the campaign down on July 28, banned the accounts involved, and passed its findings to other AI labs.
A fight over how the gap gets closed
OpenAI, Anthropic and Google have all argued publicly that distillation lets competitors catch up without matching their spending on training infrastructure and research. US agencies have separately said they've identified Chinese firms targeting American AI models for this kind of extraction.
The labs are responding with defenses built into the models themselves. Anthropic said earlier this month that new safeguards on its Claude models had cut the amount of data harvested in distillation attempts by somewhere between 90% and 99%, though it didn't say how it arrived at that figure or over what period it was measured.
Meanwhile, both Chinese models have turned in strong numbers on their own developers' benchmarks. Moonshot says Kimi K3.1 Pro beats GPT-6 Astra on a coding benchmark. DeepSeek positioned V4.1-Flash, introduced in September, as a cheaper alternative to US-built models. Neither set of results has been checked by an outside party, and neither company has commented publicly on the distillation report.
The situation leaves OpenAI and its peers with an awkward problem. The same models they sell to developers and consumers are also the easiest source of training data for anyone trying to build a competitor. Restricting access too aggressively risks annoying paying customers. Leaving it open invites exactly the kind of copying The Information says just happened.
What happens next likely depends on whether OpenAI treats this report the way it treated the July campaign, with a public accusation and account bans, or lets it pass without comment. Given the pattern with Moonshot, a response seems more likely than not.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.