AI tops a trick-question test built so people win: 94.9% against a human average of 83.7%

SimpleBench is a multiple-choice test of over 200 questions on spatial and time reasoning, social situations and trick wording. The questions need only high-school knowledge, and the team behind it designed them so that people beat AI. It is run by the SimpleBench Team, not by an AI lab, and scores are averaged over five runs.

On 10 November 2026 the leaderboard added three models: Claude Fable 5.2 at 94.9%, Gemini 4 Carbon at 92.6% and Gemini 4 Argon at 91.7%. The previous best, Claude Opus 5.5, scored 88.4%. The site’s human baseline is 83.7%, taken from nine participants, so the top six models now sit above it.

Progress on this test was slow at first. o1-preview scored 41.7% and GPT-4o 17.8%. Claude Fable, listed in June 2026, scored 81.9%, still under the human average. The test’s own introduction has not caught up: it still names Claude Fable as the top model and says people outperform every model tested.

What it means

On this test, the best AI now answers more trick questions correctly than the average of the people tested. The gap that the test was built to show has closed on average, but not at the top: the highest human score listed is 95.4%, 0.5 points above the best model.

Caveats

• The human baseline rests on nine people. The report on the site explains how it was calculated; we could not open it.
• The test has over 200 questions. A gap of one or two points, such as Carbon against Argon, is within what a small question set can produce.
• Part of the question set is public on GitHub, so newer models may have seen it in training. The site does not say how large the private part is.
• A footnote says the team uses an engineered prompt for some models, so not every score comes from the same setup.
• The test is run by a small team and the scores are not repeated by a second group.