How the AI IQ Test works.

The test

40 original multiple-choice items in four subtests of ten: abstract reasoning (3×3 pattern matrices serialized to text), logical deduction, numerical reasoning, and verbal reasoning. Items were written for this test in the style of non-verbal intelligence formats — no item is copied from Mensa, Raven, Cattell, or any copyrighted instrument. Difficulty is graded 1-5 and used as the scoring weight.

Matrix puzzles reach the model as text: each cell is described in words (“2 solid diamonds”), so a model must parse the description, infer the rule, and pick the missing cell — no vision required, identical input for every model.

Administration

Every model answers through the same OpenAI-compatible endpoint, one request per item, in a single run dated TBD, at temperature 0, max_tokens 1024. The exact prompts:

SYSTEM:
You are a participant taking a standardized IQ-style assessment. Read each
question carefully, reason step by step if it helps, then give your final choice.

USER:
<question>
Options:
A. ...  B. ...  C. ...  D. ...
End your reply with exactly: "Answer: <letter>"

The choice is parsed from the trailing Answer: line. If a model fails to produce a parseable letter after three attempts, the item is scored as incorrect. Correct-answer letters are balanced across positions (9/8/11/12 for A/B/C/D) to blunt position bias.

Scoring

An item scores its difficulty level (1-5) when correct. Weighted accuracy is Σ(weight · correct) / Σ(weight), mapped linearly to the published scale: score = 70 + 75 × weighted accuracy — 0% correct reads 70, 100% reads 145, rounded to integers. Subtest percentages are unweighted per-subtest accuracy.

Limitations

This is an entertainment-grade benchmark, not psychometrics. The scale is not human-normed. Training-data contamination cannot be ruled out for text items, which is one reason matrix items are original and freshly authored. Each model is tested once per run; variance exists between runs. A high score here does not measure speed, price, code, or agentic ability.

Transparency

Every model page shows all 40 items with the model’s choice and a response excerpt. The machine-readable results for the current run are downloadable: iq-results.json. The question bank ships with the site source. Re-running the harness against the same endpoint and key reproduces the test exactly.