Exploring Claude/GPT Knowledge Cutoffs and Pre-Training Timelines

A researcher used carefully designed quizzes and probing techniques to estimate the training timelines and knowledge cutoffs of frontier AI models like GPT-5 and Claude by testing their knowledge of niche facts and historical events. By analyzing error rates across daily facts from Wikipedia, the study revealed approximate pre-training checkpoint dates for various models and found that Anthropic's Opus models from version 4.7 onwards appear to share the same underlying pre-training checkpoint. The analysis demonstrates that hidden information about how frontier models were trained can be inferred through systematic behavioral testing, though the findings remain estimates lacking official ground truth verification.