Can an AI Have an IQ? What GPT-5.6 Benchmarks Really Measure

Does the new GPT-5.6 have an IQ? The direct answer is no — not because the model is unintelligent, but because an IQ score is a statistical comparison against other human beings, and a machine cannot stand in that comparison any more than a submarine can win a swim meet. When OpenAI launched GPT-5.6 this week with the tagline "frontier intelligence that scales with your ambition," the word "intelligence" started doing two very different jobs at once. Understanding the difference between what an IQ test measures and what an AI benchmark measures will make you a sharper reader of every AI headline from now on.

What an IQ score actually is

An IQ score is not a count of how much you know or a meter reading on some inner engine. Modern tests such as the Wechsler Adult Intelligence Scale are normed: they are administered to a large, representative sample of people, and your raw performance is converted to a score based on where you fall in that distribution. The scale is deliberately anchored so the average is 100 and the standard deviation is 15. A score of 115 means you outperformed roughly 84 percent of the norm group; 130 — the neighborhood of Mensa's admission threshold of the top 2 percent — means you outperformed about 98 percent.

Three properties follow from that design, and all three matter for the AI question:

• An IQ score is relative. It has no meaning except in comparison to the human norm group. There is no absolute unit of intelligence being measured.
• The test assumes a test-naive human. Norms are only valid if you have not seen the items before and are working with ordinary human memory, perception, and motivation.
• The score summarizes correlated abilities. Psychologists since Charles Spearman have observed that performance across very different mental tasks correlates — the statistical core called g. An IQ score is a practical summary of that pattern in humans.

It is also worth remembering what the instrument was invented for. Alfred Binet designed the first practical intelligence test in the early 1900s as a school screening tool, to identify French children who needed extra help — not to rank adults or crown geniuses. The term "intelligence quotient" came later, from psychologist William Stern's ratio of mental age to chronological age, and the ratio itself was eventually abandoned in favor of today's normed deviation scores. The point: IQ has always been a comparison of humans to humans for a practical purpose, revised repeatedly as psychologists learned its limits. It was never a universal intelligence meter waiting for new kinds of minds to step onto the scale.

Why you cannot meaningfully give an IQ test to a language model

The norm group problem

Give a model a Wechsler-style item and score it against human norms and you have already broken the instrument. The norms describe test-naive humans of a certain age distribution. A model is not in the population the test describes, so the percentile — which is the entire content of the score — refers to nothing.

The contamination problem

Large language models are trained on enormous swaths of text that include IQ-test items, puzzle books, test-prep guides, and detailed discussions of standard test formats. When a model answers a classic matrix-reasoning or vocabulary item correctly, you cannot distinguish reasoning from recall. In psychometrics this is simply an invalid administration — like scoring someone who memorized the answer key.

The jagged-abilities problem

Human abilities correlate; that is what makes a single summary score useful. Machine abilities do not follow the human pattern. A frontier model can perform graduate-level chemistry reasoning in one response and miscount the letters in a short word in the next. Researchers describe this profile as jagged: superhuman on some axes, subhuman on others, with no human-like g holding it together. A single number cannot summarize a profile that does not correlate the way human abilities do.

What AI benchmarks measure instead

The scores you see in AI launch announcements — including the coding, science, and reasoning results cited for GPT-5.6's new Sol, Terra, and Luna model tiers — come from task benchmarks: fixed collections of problems with scoring rules. They measure performance on that task under those conditions. That is genuinely useful, but it differs from an IQ score in two critical ways.

First, benchmarks saturate. When models approach the ceiling on a benchmark, labs move to harder ones, so scores are not comparable across generations the way normed IQ scores are across decades. Second, benchmarks leak. Once a benchmark is public, versions of its problems tend to seep into training data, inflating later scores — the same contamination problem, industrialized. Benchmark designers now build private test sets and novelty-focused evaluations, such as the ARC-AGI benchmark created by researcher Francois Chollet specifically to test whether a system can solve genuinely unfamiliar problems rather than recombine memorized ones. The consistent finding: models score far lower on tests built to defeat memorization than their headline results suggest.

So when a launch page says a model achieves state-of-the-art results across coding and science, the accurate translation is: it outperformed other models on specific problem sets. It is a real achievement. It is not an IQ, and no serious lab claims it is.

Why the confusion persists

Journalists and marketers reach for IQ because it is the one intelligence number the public already knows. There is also a long history of viral experiments feeding IQ test items to chatbots and reporting scores of 120, 150, or higher. These make entertaining headlines and terrible measurements, for every reason above. The deeper issue is that the word intelligence was never one thing even in humans. The Flynn effect — the well-documented rise of roughly three IQ points per decade across much of the 20th century — showed that even human scores move with environment, education, and familiarity with abstract test formats. The measurement is subtle for people. For machines, the instrument simply does not apply.

The practical takeaway for reading AI news

A three-question filter handles almost every "AI intelligence" headline:

• What task, exactly? "Beat humans at X benchmark" means a specific dataset, not general capability.
• Could the test be memorized? If the benchmark predates the model's training, suspect contamination until told otherwise.
• Who is the comparison group? "Human-level" claims depend entirely on which humans, doing what, under what conditions.

That filter is itself a small exercise in real-world intelligence: not vocabulary, not trivia, but disciplined evaluation of claims. Human intelligence, unlike a benchmark score, shows up in how you reason when the problem is unfamiliar and the stakes are real. Across 180 million total solvers of Timothy E. Parker's puzzles over 30 years and more than 80 countries, the durable observation is that this kind of applied reasoning is trainable — people get measurably better at spotting traps and structuring problems with practice. If you are curious how your own practical reasoning stacks up — the human kind, measured the way it should be, against other people — the Real World IQ test measures exactly that: judgment on realistic problems rather than abstract puzzle-book items. It is a far more honest mirror than borrowing a machine's benchmark, and considerably more useful than wondering what number a chatbot would get on a test it has already read about a thousand times. You can take the Real World IQ test in one sitting and get a normed comparison against real human performance.

FAQ

Does GPT-5.6 have an IQ score?

No. IQ scores are percentile comparisons against a normed human population, and a language model is not part of any human norm group. GPT-5.6's launch results are task-benchmark scores in areas like coding and science, which measure something different.

Why do some articles say chatbots have an IQ of 120 or 150?

Those numbers come from feeding human test items to a model and scoring them against human norms. The result is invalid twice over: the model is outside the norm population, and test items and formats appear in its training data, so recall is indistinguishable from reasoning.

What is the difference between an IQ test and an AI benchmark?

An IQ test is normed against a representative human sample and reports your standing relative to other people. An AI benchmark is a fixed problem set that reports task performance, saturates as models improve, and is vulnerable to training-data contamination.

Is there any test that measures general intelligence in AI?

Researchers have built evaluations aimed at generalization to genuinely novel problems, such as Francois Chollet's ARC-AGI. Models score much lower on these than on conventional benchmarks, and there is no consensus measure of general machine intelligence.

What does an IQ score of 130 actually mean?

On the standard scale with a mean of 100 and a standard deviation of 15, a 130 places you around the 98th percentile of the norm group — roughly the threshold Mensa uses for membership.