How AI models score on our IQ test

Results

Mean is the average estimated IQ across 5 tests. Range is the lowest and highest single test. Strongest section is the section where the model answered the largest share of puzzles correctly across all its tests.

Share of puzzles answered correctly, by section

Columns, in test order: Pattern Recognition (5 puzzles per test), Abstract Reasoning (9 puzzles per test), Spatial Reasoning (5 puzzles per test), Quantitative Reasoning (3 puzzles per test), Logical Reasoning (3 puzzles per test).

Every test

Each cell is one test's estimated IQ, with the number of puzzles answered correctly under it. The same test number means the same puzzles for every model.

The full write-up, with the data as a CSV download and a suggested citation, is the study AI vs human IQ 2026.

How we tested

Each model took the same test a person takes on IQ Octopus: 25 puzzles in five sections (Pattern Recognition 5, Abstract Reasoning 9, Spatial Reasoning 5, Quantitative Reasoning 3, Logical Reasoning 3). The puzzles are generated from a seed, so we fixed 5 seeds (1001, 1002, 1003, 1004, 1005) and gave every model the same 5 tests. That is 125 puzzles per model.

What the model saw

Each puzzle was rendered as one PNG image, the way it looks on screen: the question at the top, the puzzle below it, and the answer options underneath, each labeled with a number from 1. Black lines on a white background. For the number series puzzles we also sent the numbers and options as plain text alongside the image. Each puzzle was a separate request with no memory of the others, and the model never learned whether it was right.

The exact prompt

System prompt:

You are taking a visual reasoning test. Each question shows one puzzle as an image: the puzzle on top and numbered answer options underneath. Exactly one option is correct.

User turn: the image, then the same question a person sees above that puzzle, followed by the answer instruction:

Which tile comes next? Answer with the option number only.
Which tile completes the grid? Answer with the option number only.
Which option is the same shape, only rotated? Answer with the option number only.
Which one does not belong? Answer with the option number only.
Which number comes next? Answer with the option number only.
Which tile completes the second pair? Answer with the option number only.

The OpenAI models ran through the Codex command line tool, which takes one message. For them the system prompt, the line “The attached image is the puzzle.” and the question above were sent together as that message, with the same image attached.

Settings

• Claude Opus 5.5 (Anthropic): Claude API, adaptive thinking, effort medium.
• GPT-6-Astra (OpenAI): Codex CLI, reasoning effort medium.
• Claude Sonnet 5.5 (Anthropic): Claude API, adaptive thinking, effort high.
• Claude Haiku 5.5 (Anthropic): Claude API, adaptive thinking, effort medium.
• GPT-5.6-Terra (OpenAI): Codex CLI, reasoning effort medium.
• Temperature: provider default (these models do not accept a temperature setting).
• Each model could reason before it answered, up to 16,000 output tokens for Claude models. Only the final reply was read.
• Tests run on 2026-10-21.

Scoring, identical to people

Most replies were a single number. When a model added an explanation, we took the number on its own line at the start or the end of the reply, and otherwise the first number in the reply. A reply with no valid option number counts as a skipped puzzle, and a skipped puzzle counts as wrong, exactly as it does for a person. The answers then went through the same scorer that scores every human attempt: a Rasch model that weighs each puzzle by its difficulty and turns the pattern of right and wrong answers into an estimated IQ. See how the test works for the details. A model's score on this page is the mean of its 5 estimated IQs, rounded to a whole number the way the scorer rounds.

Limitations

• Models see a rendered image and get no time limit. People have 20 minutes for the whole test. The models had no clock, and a person sees the same picture on a screen, not a file.
• The scale was built for people. The puzzle difficulty values are set for human test takers. An estimated IQ for a model translates its right and wrong answers onto that human scale. It is not a measure of intelligence in the human sense.
• Results vary by prompt and settings. A different prompt, more reasoning time or a different image size could move these numbers. We used one fixed prompt and the settings listed above.
• We did not test every model. We tested the models we could run on 2026-10-21. Gemini models are not included because we had no access to them for this run.
• 5 tests per model is a small sample. The range column shows how much a model's score moved between tests. Treat differences of a few points as noise.