Blogger Leo is stirring up drama: they claim DeepSeek V4 Pro's official API might be routing some tough coding requests straight to Anthropic's Claude Fable 5 — all to hoover up outputs for model distillation (training a weaker model on a stronger one's answers).
Testers used OpenCode to have the model generate a 3D game. Some results looked eerily like Fable 5, with reasoning suddenly shifting風格. When they tossed in cybersecurity and biology questions, the game quality nosedived, and the knowledge range snapped back to DeepSeek's usual level.
The test exploits Fable 5's real behavior: Anthropic already routes some cybersecurity, biology, and distillation requests to Opus 4.8. So Leo suspects DeepSeek sends requests to Fable 5 first, and only falls back to DeepSeek after triggering a classifier.
But here's the catch: all we have is weird API behavior. It can't prove who actually answered, or that outputs entered the training data. Prompt blending, unannounced updates, or DeepSeek's own model routing could also explain the discrepancies.
https://twitter.com/synthwavedd/status/2078514339552628880?s=46