DeepSeek V4 Pro API Accused of Routing Requests to Claude for Distillation

On July 19, 2026, whistleblower leo raised concerns on X: DeepSeek V4 Pro's official API routes some complex programming requests to Anthropic's Claude Fable 5. The allegation was forwarded by user 0xLogicrw, who stated the purpose might be to collect outputs for model distillation—training one model using another's answers.

Testers used OpenCode to generate 3D games and found some results highly similar to Fable 5, with a sudden change in reasoning style. This phenomenon was cited as indirect evidence of routing behavior. However, current evidence only proves abnormal API behavior, not confirming the actual responder or training data usage.

The allegation suggests routing behavior may be domain-specific: when cybersecurity and biology questions were added, game quality significantly dropped. This implies DeepSeek might trigger routing only on specific complex programming tasks, not all requests.

Notably, Anthropic itself has a similar routing mechanism: it routes some cybersecurity, biology, and distillation requests to Opus 4.8. This context increases the credibility of the DeepSeek routing allegation but does not directly confirm it.

All current evidence comes from third-party tests and indirect observations. The key issue: it cannot confirm whether the API's returned content actually came from Claude Fable 5, nor whether DeepSeek used the outputs for training.

Current evidence only proves abnormal API behavior—output patterns inconsistent with expectations—but is insufficient to conclude routing or distillation. The whistleblower themselves admitted, "Unfortunately, my tasks were too simple to trigger Fable 5 calls," indicating limited testing conditions.