DeepSeek V4 is as powerful as Fable 5, and the community suspects that the API silently switched models.

According to Perceive Beating's monitoring, whistleblower blogger leo has raised concerns that the official API of DeepSeek V4 Pro will route some complex programming requests to Claude Fable 5. The purpose may be to collect outputs for model distillation, where the answers of a strong model are used to train another model.

Testers used OpenCode to have the model generate a 3D game. Some results showed a high similarity to Fable 5 and exhibited a sudden change in reasoning. After introducing network security and biology questions, the game's quality significantly declined, and the knowledge scope regressed to DeepSeek's original level.

This test leveraged Fable 5's underlying mechanism. Anthropic will route some network security, biology, and distillation requests to Opus 4.8. Testers, therefore, suspect that the requests are initially routed to Fable 5, triggering the classifier before falling back to DeepSeek.

However, the existing evidence only proves abnormal API behavior. It cannot confirm the actual responders, nor can it prove that these outputs entered the training data. A mix of prompts, undisclosed updates, or DeepSeek's own model routing could all lead to similar discrepancies.