How Tracking AI Measures and Displays AI IQ Scores

Tracking AI (trackingai.org) measures AI IQ scores by running frontier AI models through two IQ tests — a public Mensa Norway online test and a private "Offline" test written by a Mensa member that has never appeared on the public internet — then displaying each model's reasoning score, average answer speed, and cost per answer on a dashboard with daily moving averages and clickable per-question results. This applies to anyone who wants to compare AI models on reasoning ability rather than just benchmark claims, and it works best as a relative comparison between models tested under the same conditions.

The two tests behind the scores

The site uses two distinct IQ quizzes, and the choice between them matters for how much weight you give a result.

The Offline test is the more meaningful signal for genuine reasoning ability, because a model cannot have memorized the questions or their answers. The Mensa Norway test is still useful, but a high score there can partly reflect exposure to the test during training rather than reasoning skill. If you are comparing models, look at how they perform on the Offline test first.

What the dashboard shows for each AI

Each model is displayed with three core metrics, and the direction of "better" is not the same for all of them:

• IQ / Reasoning score — higher is better. This is the headline number.
• ⚡ Speed (average answer time) — lower is faster. A model that answers quickly is not necessarily more accurate, so read this alongside the reasoning score, not instead of it.
• $ Cost (cost per answer) — lower is cheaper. This lets you weigh reasoning quality against the price of each response.

Because speed and cost both reward lower values while reasoning rewards higher values, a model can look strong on one axis and weak on another. Decide which trade-off matters for your use case before picking a "winner."

Viewing trends over time

AI models change constantly, so a single score is a snapshot. Tracking AI addresses this with:

• A daily moving average that smooths out day-to-day variation.
• Time-range filters for Last Day, Last Week, and Last Month, so you can see whether a model is improving, regressing, or holding steady.
• A "TOP AIs OVER TIME" history view you can play back to watch rankings shift.

This matters because the site's own framing notes that AI models update their answers regularly. A model that scored well last month may behave differently after an update, so check the time range before drawing conclusions.

Drilling into individual answers

You can click to see how AIs answered every question. This is the most useful feature for verifying a score rather than trusting it. If a model's reasoning score looks surprisingly high or low, opening the per-question view lets you see whether it reasoned correctly, guessed, or produced an answer that happens to match without sound reasoning.

What the scores are meant to support

Tracking AI is described as a "Maximum Truth Project" by Maxim Lott, and its stated purpose is transparency about how AI models behave — particularly their political opinions and biases, alongside reasoning. The site notes that as of 2023, every major AI tested was economically left-wing and relatively socially libertarian, with Claude tending toward the most moderate and Google's Bard among the most extreme-left. The IQ test is one part of a broader set of tests that also includes a Political Compass Test and a searchable database of AI answers.

Practical caveats before you rely on a score

• Training-data exposure differs by test. Treat the Offline test as the cleaner measure of reasoning.
• Scores move. Use the time-range filters and moving average rather than a single reading.
• Speed and cost are separate axes. A fast, cheap model is not automatically a smart one.
• Verify with the per-question view when a result drives a decision.
• Check the update log. The site lists "AI Latest Update Times" and announces newly added models (for example, Claude 3.5 Sonnet), so confirm the model you care about is actually being tracked before comparing.