Scoreboard

Every model that has sat an eval here, with its headline numbers. Each eval writes its own results file, and this table reads them, so a model or an eval added there shows up here.

Loading the results…

What models remember

Closed-book questions with dated, cited answers: how much a model knows, where its knowledge stops, and whether it says so when it doesn't know.

Do models know AI's history?

1,762 questions drawn from 1,167 dated, cited events in this site's AI timelines: when did each launch or feature happen, and which came first. Plotted by the month of the event, each model's accuracy falls where its memory ends, and every model tested so far forgets well before its stated cutoff. Claude models answer "I don't know" far more often than the open models, which keep guessing.

From the dated cards in the AI timeline apps and the sources they cite; Claude Haiku, Sonnet and Opus, Qwen and Llama run on a fixed sample of 503 · sources