Probing frontier models to work out their real training cutoffs

Shrivu Shankar ran a set of experiments to estimate when the pre-training data behind frontier models actually ends. By his measurements, Anthropic's Opus 4.7 and later appear to share a single checkpoint with an effective knowledge cutoff around late December 2025, while OpenAI's GPT-5.6 comes from a separate checkpoint that finished near late February 2026. Opus 5 is the odd one: its documentation lists a May 2026 cutoff, but it behaves like a model that stopped learning in January 2026.

The method is simple to repeat. Shankar gave each model eight-way multiple-choice quizzes on daily facts drawn from Wikipedia to find where the error rate jumps, asked each model for the current date, and ran 50 "what model are you?" prompts to infer what went into the training mix. One quirk stood out: Sonnet 5 often identifies itself as GPT-4, a hint that ChatGPT-generated text may have leaked into its training data, while OpenAI's own models never claim a competing identity.

Why it matters

If you build on these models, the cutoff printed in the documentation may not match what the model actually knows. When your application depends on recent facts, test the boundary yourself instead of trusting the label.