We can learn hidden facts about frontier models by "probing" them with curated requests. By scoring models on niche facts, we can approximate parameters or estimate training timelines. Model training typically involves three stages: pre-training on massive general-purpose data, fine-tuning on domain-specific "textbook quality" data, and post-training to develop an assistant persona.

By analyzing error rate timelines through Wikipedia daily-fact quizzes, we can estimate when a model loses signal from its training dataset. For example, Anthropic models like Opus 4.7 onwards appear to stem from a training run cutting off around late December 2025. Meanwhile, Open AI's GPT-5.6 family seems to originate from a checkpoint finished around late February 2026.

Probing also reveals how labs train on past-model outputs. Vertical bands in data suggest models are trained on user chat sessions from platforms like ChatGPT.com and Claude.ai. Interestingly, while OpenAI models rarely misidentify, Anthropic's Sonnet 5 may occasionally self-identify as GPT-4, suggesting older ChatGPT sessions might be present in the training mixture.