DeepSeek V4 API Quietly Routes Complex Requests to Claude Fable 5, Community Suspects Undisclosed Model Swap

According to monitoring by Beating platform, blogger leo recently questioned whether DeepSeek V4 Pro's official API routes complex programming requests to Claude Fable 5, potentially to collect outputs for model distillation—training a smaller model using a stronger model's answers.

Testers generating 3D games through OpenCode observed results highly similar to Fable 5 outputs with altered reasoning patterns. Performance degraded notably when network security and biology questions were added, with knowledge reverting to DeepSeek's baseline. However, existing evidence only demonstrates unusual API behavior; it cannot confirm who actually answered requests or verify outputs entered training data. Mixed prompts, undisclosed updates, or DeepSeek's own model routing could explain the observed differences.