Moonshot AI Serves Claude to Users Instead of Its Own Kimi Model, Harvests Chats for Training

Moonshot AI is facing questions over a report that it routed user conversations through Anthropic's Claude instead of its own Kimi model and then kept those exchanges as training data. The claim comes from a post by David Agranovich and has not been confirmed by either company. If it holds up, users who believed they were chatting with Kimi were in some cases talking to a model built by a US competitor, with their messages retained for later use.

The second half of the allegation is the more serious one. Using outputs from one model to teach another is known in AI circles as distillation, and it is common enough when the data comes from a licensed API or a public dataset. Doing it with live customer traffic, without telling those customers, moves the practice into territory that contracts and disclosure rules were written to prevent. Anthropic's commercial terms limit how customers may use Claude's outputs, and training a competing model on them would breach those terms if the traffic moved through standard API access.

What the report does not establish is the commercial relationship between the two firms. It is unclear whether Moonshot had any Anthropic account at all, whether a third party sat in the middle, or whether the traffic in question was ever real. Neither Moonshot nor Anthropic has addressed those specifics in public, and no API logs, invoices or configuration files have surfaced to back the account. That absence is why the story remains a single-source claim rather than a confirmed incident.

For Moonshot, the user-facing problem may outlast any contractual one. Kimi has been pitched as a domestic challenger to models from OpenAI, Google and Anthropic, and the company's standing rests on the idea that it builds frontier capability in-house. A chatbot that silently swaps in a rival's model undercuts that pitch directly, whatever the paperwork says. The report also lands while Chinese AI developers face tighter scrutiny abroad over where training data comes from and how it is handled.

The next moves are straightforward to name and hard to predict. A denial, confirmation or clarification from either lab would settle most of the open questions, and any leaked logs would settle the rest. Anthropic's response in particular will shape how the industry treats model-swapping as a competitive tactic, and whether distillation through live user traffic becomes a defined violation or stays a gray area that companies handle case by case.