Bloggers Rebut Rumors That Chinese Models Secretly Route to Claude

On September 11, in response to rumors that "Chinese models secretly route to Claude," several bloggers pushed back. op7418 and orange offered two core arguments: first, these models are open source, so their benchmark results must be reproducible by others once released, making covert routing easy to catch; second, the cost of calling Claude at scale would be unsustainable—the business logic simply doesn't hold.

Confirmed

• op7418 speculated that the "Claude getting dumber" phenomenon some users encountered may stem from API relay services: these relays claim to forward Claude but actually mix in outputs from Kimi or DeepSeek, with only a small share of genuine Claude.
• 歸藏, discussing model distillation, said distillation is definitely worth doing, but "routing a model to users" serves an unclear purpose. He rejected the claim that it's "to get users to generate distillation data for vendors," arguing that most user-generated data is of very poor quality and fundamentally unsuitable as a distillation data source—making such an approach completely uncontrollable.

Why it matters

• The incident touches on the credibility of Chinese models and the trustworthiness of benchmarks; open-source reproducibility was seen by bloggers as the key mechanism for debunking cheating rumors.
• The relay-service adulteration theory serves as a reminder: "Claude getting dumber" may not be official degradation—channel quality could be a major variable.

2026-09-11 ~ 2026-09-11 · 5 related posts

• Are Chinese models secretly routing to Claude? A cost and reproducibility rebuttal — op7418 · 2026-09-11
• Dev speculates API resellers may pass off Kimi or DeepSeek outputs as Claude — op7418 · 2026-09-11
• [source] Are Chinese Models Secretly Routing to Claude? Why the Rumor Doesn't Hold Up — oran_ge · 2026-09-11
• AI commentator argues user-generated data is too noisy to be reliable distillation source — op7418 · 2026-09-11

1 near-duplicate retellings: op7418