Intelligence, in the details.

A benchmark is an experiment.

SimAgents brings published research and interactive results into one place. You can trace a ranking back to its methods, inspect model settings, and compare outcomes within the same task.

The next step is participation: experiments where people receive the information and choices a tested model would face. Originality explores useful, unusual ideas. Other Minds explores whether a participant can learn how two anonymous AIs differ.

What a ranking can tell you.

Each benchmark measures something specific. Writing compares stories by asking judges to choose between pairs. Debate compares arguments over two debates, with each model arguing both sides. Extended Connections tests word associations with added distractions and gives progressively more credit for solving more of a puzzle.

There is no combined “best AI” score here. Different release dates, judges, model settings, and groups of models compared matter. A small rank difference is not proof of a reliable performance difference; look at the uncertainty ranges and the comparisons behind the score.

People and models, carefully compared.

Visitor attempts, practice sessions, official model results, and carefully controlled human studies are kept separate. Fair comparisons need the same tasks, information, limits, and scoring. Preview drafts are for your own practice and are not used in research.

Start with the sources.