[DeepSeek V4 as powerful as Fable 5, community suspects API secretly switched models] According to monitoring by Beating, whistleblower blogger Leo questioned whether the official API of DeepSeek V4 Pro redirects certain complex programming requests to Claude Fable 5. The purpose might be to collect outputs for model distillation, which involves training another model using answers from a stronger model. Testers used OpenCode to have the model generate 3D games. Some results were highly similar to Fable 5, and the reasoning methods suddenly changed. When cybersecurity and biological questions were added, the game quality significantly declined, and the knowledge scope reverted to DeepSeek's original level. This test leveraged Fable 5's actual mechanisms. Anthropic reportedly redirects some cybersecurity, biological, and distillation requests to Opus 4.8. Testers thus suspect that requests are first routed to Fable 5, triggering a classifier before falling back to DeepSeek. However, current evidence only proves abnormal API behavior. It cannot confirm the actual responder nor prove that these outputs were used in training data. Mixed prompts, undisclosed updates, or DeepSeek's own model routing could all cause similar discrepancies. [Original link]