DeepSeek V4 is reportedly so good that the community is crying foul. A leaker known as leo claims that DeepSeek V4 Pro's official API is quietly routing some complex coding requests to Anthropic's Claude Fable 5 — ostensibly to collect outputs for model distillation (training a weaker model on a stronger one's answers).
Testers using OpenCode to generate 3D games found that some results were eerily similar to Fable 5's output, and the reasoning style suddenly shifted mid-conversation. But when they threw in cybersecurity and biology questions, the quality dropped sharply, and the knowledge scope reverted to DeepSeek's original level. The test deliberately exploited Fable 5's...