DeepSeek V4 Pro API Suspected of Routing to Claude Fable 5
Leaker blogger Leo questions whether the official API of DeepSeek V4 Pro routes complex programming requests to Claude Fable 5 for collecting outputs for model distillation.
Testing during the generation of 3D games showed results highly similar to Fable 5, with a sudden change in inference method, and quality dropped back to original levels after adding cybersecurity and biological issues.
Anthropic has a similar mechanism for routing sensitive requests to Opus 4.8, raising suspicions that requests are first classified by Fable 5 before being reverted.
Current evidence only shows abnormal API behavior, unable to confirm the actual output source or its use for training; mixed prompts or routing updates could also explain the differences.
AI companies' routing and distillation practices are exposed, with DeepSeek facing reputational pressure, while strong models like Anthropic benefit from output demand, leading capital to concentrate on transparent routing and anti-distillation technologies.
Source: Public Information
ABAB AI Insight
DeepSeek has previously rapidly iterated open-source and commercial models, and this questioning continues the industry's long-standing controversy over closed-source API distillation, similar to the exposure of Anthropic's own routing mechanism, where testing methods utilize sensitive topics to trigger classifiers.
On the capital path, routing resources are used to collect high-quality outputs for training the next generation of models, motivated by low-cost capability enhancement, strategically hiding backend implementations through APIs to achieve competitive advantages.
Similar cases include multiple companies being accused of using strong model distillation; currently, cutting-edge AI is in a stage of an arms race regarding routing and anti-distillation.
Essentially, this is about capital concentration: strong model outputs become scarce training data, with the mechanism of API routing reducing self-training costs, driving capital towards companies with top models, enhancing their pricing power and technical barriers in next-generation training, while accelerating industry discussions on transparency and ethics regulation.
ABAB News · Cognitive Laws
Complex request routing, distillation shortcuts
Sensitive topics are the litmus test for model routing
Strong model outputs are the most expensive data