Anthropic alleges Moonshot routed to Claude

Diving deeper into

Moonshot AI

Company Report
Anthropic accused Moonshot in September 2026 of routing requests to Claude and using its responses to train Moonshot's software,
Analyzed 9 sources

The key implication is that Moonshot’s biggest risk is no longer just model quality, it is trust in how Kimi is built and what system is actually answering users. Anthropic said in a September 10, 2026 threat report that Moonshot silently forwarded some Kimi requests to Claude, returned those answers as Kimi output, and collected the exchanges for training. That shifts the issue from ordinary model competition to provenance, customer disclosure, and compliance risk ahead of Moonshot’s planned Hong Kong IPO.1 2 3 4

  • This is more serious than simple distillation. Distillation usually means hitting a rival API to harvest outputs offline. Anthropic alleged live request routing, where an end user thinks they are talking to Kimi but some prompts are actually answered by Claude behind the scenes. That creates product identity risk and possible user data handling concerns at the same time.1 2 5
  • The timing matters because Moonshot is trying to turn K3 into a licensed model business. Its commercial stance already asks large hosts to negotiate for access and revenue share. Enterprise buyers and cloud partners can accept aggressive economics if a model is unique, but they are much less likely to accept provenance controversy that could bring audits, contract friction, or regulator attention.4 6
  • There is already precedent for model companies getting squeezed when they depend on outside labs. In AI applications like Harvey, GC AI, and Tasklet, reliance on Anthropic flows through into pricing, reliability, and platform risk. The Moonshot allegation is a more extreme version, where dependency is not just commercial, it reaches into the origin of the model behavior itself.7 8 9

Going forward, frontier model vendors will sell not just intelligence but clean lineage. Labs that can prove where outputs came from, what data touched customer prompts, and how training sets were assembled will have an easier path to enterprise contracts, cloud distribution, and public market scrutiny. For Moonshot, that makes governance and disclosure almost as strategic as benchmark performance.1 3 4 6