Testing Model Knowledge Cutoffs with Events

giffmana · x · 2026-07-11

This post introduces a simple **model knowledge cutoff benchmark**: the author fabricates a batch of "unexpected events" for each month and asks the models questions about them, using their awareness to reverse-engineer their knowledge update boundaries. The original post also makes an observation: models from **OpenAI and Anthropic** seem to be among the few maintaining up-to-date knowledge. Other vendors, including some Chinese AI labs, are generally at least **12 months** behind. The author notes that the highlighted blocks in the chart represent their inferred actual cutoffs.

Original post →

More from Research

Research channel →