AI scores 151 on the Mensa Norway IQ test, surpassing 99.97% of humans.

AI has completely blown past human IQ test limits.

Mensa Norway IQ test: 35 extremely difficult visual reasoning questions, 25-minute time limit.

Claude Fable 5.1, along with GPT-6 Astra capable of image-based responses, entered the exam and answered every question correctly, achieving the theoretical maximum score of 151.

And it’s the seventh consecutive perfect score. 151 is the top score on the test, but no one knows where their own ceiling lies.

Among humans, nearly 98% of people worldwide can’t even reach the 130-point threshold for Mensa membership—we’re most likely part of the vast majority who are outclassed.

If you don’t believe it, try tackling question 33 on this test—it’s considered hellish difficulty. The answer is at the end of Section One.

Mensa Norway Test Question 33

Don’t get discouraged if you’re confused. This question stumps not only people but also most AI systems.

What makes this test so tough is that it doesn’t test any prior knowledge—it just throws you a bunch of confusing, nonsensical diagrams and sees if you can untangle the pattern from the chaos.

Psychologists refer to this ability to identify patterns on the spot as "fluid intelligence," which is also humanity's final defense—our claimed edge over machines.

Now, this line of defense has been completely obliterated by AI.

When you open the TrackingAI leaderboard, the first thing you see is a bell curve.

That’s the distribution of human IQ, with the vast majority clustered around the peak near 100.

In the past, AI avatars were scattered sparsely in the low-density area on the left side of the curve.

Now, they are all squeezing onto the lone tail at the far right of the curve, their avatars overlapping. Next to them stands a vertical line labeled “Highest Possible Score in This Test (Temporary).”

On TrackingAI, the Mensa Norway scores for each AI, with the rightmost vertical line representing the perfect score for this test.

This ranking was compiled by American journalist Maxim Lott.

He acts like a strict proctor, administering the Mensa Norway test weekly to over 30 leading AI models—some answer based on textual descriptions, while others analyze the original images directly. Scores are calculated as the average of the last seven attempts, so there’s no chance of slipping onto the leaderboard through a single stroke of luck.

The Mensa Norway test he used was originally designed for humans. When people take this test, the system reports a maximum score of 145 and does not differentiate higher scores.

TrackingAI scores AI based on the number of correct answers, with 35 correct answers equating to a score of 151—exceeding the upper limit of the IQ scale designed for humans by Mensa.

However, a ceiling of 145 is just the limit of this test—the intelligence of the population has no ceiling.

We estimated based on the standard distribution of IQ scores that about 1 in 3,000 people has a score of 151.

Among the global population of over 8 billion, only about 2.7 million people score above 151—roughly three thousand in a city of ten million.

The Mensa entrance threshold is 130 points; these two AI systems have surpassed it by a full 21 points.

And 151 is merely the ceiling of this test—if the scale extended further to 160, only around 200,000 people worldwide would remain.

AI is now at the very top of the human intelligence pyramid.

And far more than just these two monsters can make it into this small group.

On TrackingAI’s leaderboard, GPT-5.6 Sol and Gemini 3.1 Pro both scored 145, with numerous models now exceeding 140 on the public benchmark, and domestic models have entered the top tier.

So many models are clustered in the high-scoring range, with differences emerging only on the final few challenging questions.

According to TrackingAI, nearly any AI could score on the first 10 questions, but by the end of the test, only five AIs got question 35 right, and only two managed to solve question 33.

According to Lott’s scoring, getting one question wrong out of 35 results in a score of only 148, so the two who answered question 33 correctly are essentially the two perfect scorers.

The first question mentioned that the answer would be posted here; the correct option is E.

The number of AIs that answered each question correctly; orange is the Mensa Norway test, and only 2 remain for question 33.

A perfect score is already alarming, but consider that just a few years ago, AI was still a student with severe subject weaknesses.

In March 2023, a psychological assessment expert in Finland administered a standardized Wechsler Adult Intelligence Scale to ChatGPT, testing only verbal subtests such as vocabulary and general knowledge.

ChatGPT achieved a language IQ of 155, outperforming 99.9% of humans.

It immediately freezes when faced with a trick question like, "What's the name of Sebastian's child's father?"

The non-verbal aspects are impossible to measure; in the words of this expert, ChatGPT "has no eyes, ears, or hands."

It’s clear that language and knowledge have always been AI’s strongest subjects, while graphical reasoning remains its biggest weakness.

And the Mensa Norway test specifically assesses graphical reasoning.

This type of matrix puzzle was first designed by British psychologist John Raven in 1936, and psychologists generally regard it as the test format most capable of measuring a person’s overall intelligence.

AI has been tackling this tough problem for 60 years, since the first program to solve geometric analogy problems was developed at MIT in 1962.

We have compiled the timeline of AI solving IQ test questions based on publicly available information.

Until early 2024, Gemini Advanced could seriously fabricate a string of equations like "A+B=C" for the simplest question on the Mensa test, even though the question contained no equations at all.

In February 2024, Gemini Advanced made up a series of equations when answering question 1 from the Mensa test.

Lott read each question aloud to the AI, but the results weren’t much better. Some AIs guessed only six correctly, scoring a dismal 64—roughly equivalent to randomly filling in answer sheets with eyes closed. Even the best, Claude-3, barely surpassed the human average.

At the time, Lott asserted that AI did not yet possess human-like general intelligence. Few expected that just over six months later, the turning point would arrive.

In September 2024, OpenAI's o1 learned to think before speaking, engaging in internal reasoning, step-by-step trial and error, and self-checking—causing its score to surge past 120.

Lott, who claimed just over six months ago that AI wasn’t viable, has now written the headline: “A Major Breakthrough in AI Intelligence.”

From then on, thinking before answering became a standard feature of all flagship models, with scores steadily rising; in April this year, the leaderboard saw the first-ever score of 151.

We have compiled the data from TrackingAI and Maxim Lott; each point represents the highest recorded value at that time.

And so, a student who had been struggling in one subject for 60 years improved from a guesswork score of 64 to a perfect score of 151 in just two and a half years.

According to Lott’s statistics, from May 2024 to October 2025, the IQ scores of leading AI models increased by an average of 2.5 points per month.

In comparison, human growth is remarkably slow, and there is existing psychological data to support this.

Throughout the 20th century, as nutrition and education improved, human IQ test scores rose by about 3 points per decade—a phenomenon known as the Flynn Effect.

Compared to this, humans took over 30 years to climb 10 points, while AI achieved the same gain in just four months.

It’s rising so fast—any normal person’s first reaction would be, did it secretly memorize the answers?

Lott also doubted it back then.

After all, the Mensa Norway test has been online for many years, and both the questions and answers have long been fully exposed by netizens.

Even TrackingAI itself has publicly released the text of all 35 questions along with the correct answers, so the AI likely encountered them during training.

The text version of the Mensa Norway questions released by TrackingAI, with the correct answers listed in the right column.

To rule out this possibility, in May 2024, he specifically enlisted a Mensa member to create an entirely new set of questions from scratch.

This set of questions was only used once for reader trials to establish scoring standards and has never been published online or included in any AI training data.

After switching to the new test set, the practice effect clearly emerged. Even today, some models perform drastically worse on this confidential test set, scoring about 20 points lower than on the public one.

Even with a completely new test paper, the top model still scores nearly perfect.

The secret test contains only 16 questions, with a maximum possible score of approximately 136; OpenAI’s GPT-5.6 Terra image version has already reached 135.

Fable 5.1 and Astra also both scored 130, alongside a domestic model standing side by side.

The score distribution of the confidential exam; this exam has never been posted online, and the vertical line represents its perfect score.

Last October, Lott confidently set a goal, estimating that AI would need at least two more years to achieve a perfect score on the visual version of this exam, promising to return on Halloween 2027 to check the results.

Less than a year later, GPT-5.6 Terra surged ahead across the finish line, over a year ahead of his prediction.

Since guessing the answers doesn't make sense, the only remaining conclusion is that the AI, which was still guessing blindly two and a half years ago, has truly become smarter.

The exam paper was filled to capacity, so the question setters had to move on to even harder ones.

Therefore, what TrackingAI marks next to the perfect score line is the "Highest Possible Score (Temporary)."

Last year, Lott planned to introduce a new set of more challenging questions in the score range above 130 this year, in order to re-establish a gap among high-scoring AIs.

Of course, some people also disagree.

François Chollet, the creator of Keras, designed ARC-AGI, a set of reasoning tests specifically crafted to challenge AI.

He recently stated that if intelligence is defined as the efficiency of turning experience into ability—meaning who learns faster from the same number of examples—current AI is still about six orders of magnitude behind humans, roughly a million times slower.

Yet Chollet himself sets the finish line for AGI at problem-setting.

In February this year, he revealed that ARC-4 is working on something big, set to launch early next year, with subsequent generations—6th and 7th—planned to follow.

In his own words, the key is to keep coming up with new problems until there are no more tasks that humans can do but AI cannot.

At the time, he estimated AGI would arrive around 2030.

At the beginning of this month, someone asked him whether his 2030 prediction still held, and he revised his statement, saying it would happen even earlier due to faster progress than he had anticipated.

Someone asked Chollet whether his judgment that AGI will be achieved by 2030 still stands.

The IQ test was created in 1905. For 120 years, humans have readily used this measuring tool to rank one another, from school tracking to military selection.

Behind this ruler lies an assumption that humans have almost never questioned: only humans can be measured, and only humans can reach the highest mark.

The ruler still exists, but the premise has collapsed. Something that isn’t even human crawled from start to finish along the markings, and not even the topmost line could stop it.

Lott can create even more challenging tests, and Chollet can follow up with the next generation of ARC.

With AI improving from blind guessing to perfection in just two and a half years, no one dares to guarantee how many problems remain unsolvable by AI in the human brain.

On the day when the human knowledge base is completely exhausted, we will need to reconsider not only how intelligent AI truly is, but also what remains of humanity once we strip away the facade of "intelligence."

Reference materials:

https://x.com/aisafetymemes/status/2103369359481852116

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.