GPT-5.6’s 136 IQ Score Is Real. ‘Smarter Than 99% of Humans’ Doesn’t Follow.

ByGeoff Dyers
gpt-5.6 iq score: An IQ-style matrix puzzle sheet with pencil and answer choices, suggesting benchmark scoring and test interpretation.

David Scott Patterson, an investor who posts frequently about AI progress, wrote on X that “GPT-5.6 gets an initial IQ score of 136, which is smarter than 99% of humans”, and that OpenAI’s newest model is “the first model to score over 130”. Marc Andreessen replied with a single word — “Interesting.” — and the number travelled from there. It comes from TrackingAI, a site that runs frontier models through IQ-style tests and plots them on the same 15-point-per-standard-deviation scale used to report human IQ, where 136 does sit around the 99th percentile.

The score is real, and “smarter than 99% of humans” still does not follow from it. The reason has little to do with the model and everything to do with which test produced the number and how that test was administered. TrackingAI says it runs two different tests — the public Mensa Norway online test and an offline test of about 100 pattern-recognition items “written by a Mensa member” and “not published on the internet” — and that non-vision models get the items verbalized instead of seeing the original image. Both facts change what a 136 is evidence of, and neither is visible in the tweet.

Offline 136, not Mensa Norway

Only one of the two tests makes “the first model to score over 130” a true statement, and it is the offline one. TrackingAI describes that test on its methodology page as roughly 100 pattern-recognition questions, written by a Mensa member, with the items kept off the public internet.

Models cleared 130 on the public Mensa Norway test a long time ago. Contemporaneous reporting said OpenAI’s o3 scored 136 on the Mensa Norway test in April 2025, and Forbes later reported that GPT-5.2 scored 147 on Mensa Norway after OpenAI released GPT-5.2 on December 11, 2025. One caveat belongs on the record: Patterson’s post does not say which of the two tests produced the 136, and TrackingAI publishes its per-model scores in a live chart rather than in the page text. The offline-test reading is an inference from “first model over 130”, not something TrackingAI stated.

The public and offline tests also have different statuses. Mensa Norway’s own page calls its version an informal online test that can indicate IQ level and uses the same type of tasks as its authorized test. Mensa International’s FAQ likewise distinguishes practice-style online tests from formal supervised assessments. So, even before AI enters the picture, a score on Mensa Norway’s online test is not the same thing as a full human intelligence assessment.

That makes the strongest objection to the offline 136 surprisingly narrow: it is probably not answer memorization. TrackingAI’s whole point in building the offline set is that the items are not on the public internet. If the benchmark is working as described, “the model saw these exact answers in training” is the lazy objection, and not the good one.

The better objection is about format exposure, not item leakage. Matrix reasoning, sequence completion, and Raven-style analogies are all over the public record in textbooks, worked examples, psychometrics papers, practice sites, and tutoring material. A model can avoid seeing these exact items and still absorb a large library of how such items are constructed and solved. That is still evidence of capability. It is just not the same thing as showing that the model possesses the human construct those tests were designed to estimate. A psychometrics-focused critique from QuantUX makes exactly that point: success on human IQ items can reflect learned statistical associations and test-taking strategies rather than the same underlying construct psychologists intend to measure in people.

Verbalizing visual matrices changes the task

TrackingAI says it verbalizes IQ-test items for non-vision models, and that changes what is being measured. Its methodology page says that for non-vision models, the IQ tests are verbalized, while vision models are shown the image directly.

That is not a clerical detail. A visual matrix problem and a text description of that matrix are not the same task. The first asks a system to parse shapes, spatial relations, transformations, and missing elements directly from an image. The second asks it to reason over a linguistic encoding of those relations. For a language model, that can be the difference between “see the pattern” and “read a structured hint about the pattern.”

Published evidence backs that distinction. A 2025 peer-reviewed comparative study in Intelligence found that large language models performed substantially better on verbal than on visual IQ-style tasks, and discussed prior results showing that multimodal models such as GPT-4V still struggled on more complex Raven-style visual reasoning. The paper does not prove that every LLM is bad at every visual task. It does support the narrower, load-bearing claim here: verbal and visual IQ-style tests do not probe the same abilities equally well in LLMs.

That means a single human-style IQ number can hide a lot. A verbalized matrix score may tell you that a model is good at symbolic pattern completion expressed in text. It tells you less cleanly that the model has matched human visuospatial reasoning. On a benchmark like ARC-AGI-3’s human baseline, readers already accept that test format matters. IQ-style reporting should get the same scrutiny.

“The model can do well on the benchmark and still not be measuring the same thing the human test was built to measure.”

The model can do well on the benchmark and still not be measuring the same thing the human test was built to measure.

Human IQ testing is contested before AI enters the picture

Human IQ testing was already a contested measurement problem before anyone tried to map frontier models onto it. IQ tests are useful, widely studied instruments. They are also not a clean, uncontested meter of “intelligence” in the singular.

The most defensible version of the human-side critique is not that IQ tests are meaningless. It is that construct validity is debated, norming matters, and online samples drift. The VentureBeat report on TrackingAI notes both the appeal of a human-readable number and the methodology concerns around opacity and interpretation. Mensa Norway’s own page frames its test as informal, and Mensa International separates online tests from formal supervised evaluation. Those are ordinary caveats in human psychometrics. They become bigger once the subject is a model answering reformatted items.

A human IQ score also sits on a normed scale built around populations and test administration conditions. “136” is a percentile claim only within a particular scaling framework. On the common 15-point standard-deviation scale, 136 is indeed around the 99th percentile. The arithmetic is not the weak part. The word smarter is. A model can hit a high score on one benchmark family without that justifying a broad, human-comparative statement about general intelligence.

This is the point many hot takes skip. We are grading machines on instruments humans still argue about. The AI result inherits the old disputes, what the construct is, how stable the norms are, what the subtests measure, and adds new ones about training exposure, format conversion, and modality mismatch.

That does not leave nothing. Rising scores across the same benchmark family are still evidence of real improvement. They show that frontier models are getting better at the kinds of pattern-reasoning tasks these tests contain. The mistake is turning that narrower result into a sweeping statement that the model is simply smarter than almost every human.

The practical rule for the next viral IQ post is plain: ask which test, administered how, and against what norm. If the answer is “TrackingAI’s offline test, with verbalized items for non-vision models,” then the number is interesting, but it is not self-explanatory.

Key Takeaways

  • GPT-5.6’s reported 136 appears to refer to TrackingAI’s offline IQ-style test, not the public Mensa Norway online test.
  • TrackingAI says it runs two different tests and verbalizes visual items for non-vision models, which changes the task being measured.
  • The strongest critique is not exact-answer memorization but prior exposure to matrix-style formats and solving strategies documented across public training data.
  • A peer-reviewed 2025 study found LLMs perform better on verbal than visual IQ-style tasks, which makes verbalized scores harder to compare with human visual reasoning.
  • A rising IQ-style score is evidence of improving pattern-reasoning performance, not clean proof that a model is “smarter than 99% of humans.”

Further Reading