AI Scores on the Mensa Norway Test: What 151 Means and What It Does Not

On 28 September 2026 a Chinese technology outlet, Xinzhiyuan (AI Era), republished by 36Kr, reported that an AI pairing of "Claude Fable 5.1" with "GPT-6 Astra" had "hit the theoretical maximum score of 151" on the Mensa Norway IQ test, with full marks in seven consecutive recent attempts. The numbers come from TrackingAI, the public leaderboard run by the journalist Maxim Lott that gives AI models a Mensa Norway test and an unpublished "offline" test each week. TrackingAI publishes its raw score log as an open file, so the claim can be checked. We read the log as it stood on 7 October 2026. The short version: 151 is real as a number and thin as a measurement, and the log does not show seven perfect sittings in a row.

How does TrackingAI turn right answers into an IQ?

The Mensa Norway test is 35 visual-pattern puzzles with a 25-minute limit, taken online. Mensa Norway itself reports a human result as an IQ "between 85 and 145, where 100 is the population average" with a standard deviation of 15, and says on the same page that the test is not a substitute for professional intelligence tests. TrackingAI does not stop at 145. The chart code on trackingai.org converts each raw score with one line of arithmetic: IQ = 63.5 + 3 x (correct answers – 5.833), rounded to a whole number.

That makes the scale easy to read and easy to over-read. Each correct answer is worth exactly 3 points. Eighteen right answers give 100, 28 give 130 and about 31 give 140. Zero right answers still give 46, because the line has an intercept and no floor. The top is 35 right answers, which gives 151.0. A model that answers all 35 puzzles correctly has not been measured at 151; it has been measured as "at least as high as this test can say". That is the same ceiling problem we describe for human tests in why scores stop near 160.

• 33 correct = IQ 145; 34 correct = IQ 148; 35 correct = IQ 151
• 28 correct = IQ 130; 18 correct = IQ 100; 12 correct = IQ 82
• Mensa Norway's own human scale tops out at 145, so any TrackingAI figure above that is an extension of the line, not a score a human can receive on the same page

What does the score log actually show?

The file is TrackingAI's daily_logs.csv. On 7 October 2026 it held 5,775 rows, covering 9 May 2024 to 7 October 2026. Of those, 1,906 are Mensa Norway sittings across 84 named entries, each scored out of 35, and 1,964 are offline-test sittings scored out of 16. Many entries come in two forms, a text-only form that is given the puzzles described in words and a "Vision" form that is shown the images, so the same underlying model can appear twice.

The climb is steep. In the first weeks, up to 1 June 2024, the 94 Mensa Norway sittings averaged 12.0 correct out of 35, an IQ of 82 on TrackingAI's line, and the best single sitting was 18 (IQ 100). The 88 sittings logged in September and October 2026 averaged 30.3 correct (IQ 137), with a median of 32 (IQ 142). The best "average of the last seven sittings", the figure TrackingAI charts, stood at about 94 at the end of May 2024, 121 at the end of September 2024, 133 in June 2025, 143 in September 2025 and 145 in June 2026.

Perfect scores are rare. Only 11 of the 1,906 sittings, 0.6%, scored 35 of 35. The first was Grok-4 on 1 April 2026, and the others, in date order, were Gemini Pro (22 April), Grok-4 (18 June), Grok-4-5 (23 July), GPT-Sol-Ultra (3 August), Gemini Pro (24 August), Claude-Fable (3 September), GPT-Sol-Ultra-Vision (3 September), GPT-Astra-Ultra-Vision (10 September), GPT-Terra-Ultra (30 September) and Claude-5-Opus-Vision (3 October). That is nine entries, and no entry has two perfect sittings in a row.

Where did the 151 headline come from?

The 36Kr report says the figure is the average of the last seven attempts. That rule is where the headline thins out. On 28 September 2026, the day of the report, the highest average-of-seven on the Mensa Norway chart belonged to GPT-Astra-Ultra-Vision, and that average was 35.0, or 151, because the entry had exactly one Mensa Norway sitting on record, on 10 September. An average of the last seven sittings is the average of the last one when only one exists.

The entry has since been sat twice more, on 30 September and 5 October, and scored 33 both times. Its average of three is 33.7, which is an IQ of 147 on TrackingAI's line. The runner-up, Claude-Fable, has two sittings (32 and 35, average 33.5, IQ 147). Models with seven or more sittings, which are a steadier basis, sit at 145: Grok-4 (33.0), GPT-Sol-Ultra (33.0), GPT-5-Pro-Vision (33.0) and Gemini Pro (32.9). We could not find the seven-in-a-row run in the log, and the report does not say how a "pairing" of two models was scored. It may describe a set-up that TrackingAI does not publish. Treat the claim as unverified.

What does the unpublished offline test add?

TrackingAI gives every model a second test, written by a Mensa member and never posted online, so that it cannot be in training data. Lott described it in a post of 6 May 2024: 16 new items were normed against Mensa Norway items with 35 usable responses from blog readers, two items that behaved badly were dropped, and a third with two correct answers was kept. That left 14 scored items, which is why the log's "valid score" never exceeds 14.

Fourteen items make a coarse ruler. The chart code gives IQ = 63.5 + 3 x 0.8823 x (valid score – 3) x (35 / 14), so one item is worth about 6.6 points, a score of 13 gives 130, and the maximum of 14 gives 136. In the log, 23 offline sittings scored 14 of 14, the first on 5 September 2025. For the 23 entries active since 1 August 2026 with at least three sittings on each test, the public test ran on average 13.9 points above the offline one (median 14.0), which echoes the 13.5-point gap we reported on 5 September in our guide to the ChatGPT IQ score. The gap reached about 20 points for entries such as Claude-5-Sonnet, Grok-4-5-Vision and deepseek-v3. The offline test was normed on only 35 people, and Lott reported a correlation of R-squared .45 between the two item sets, so it is a useful check on memorising rather than a precise second IQ.

Does an AI "IQ" mean what a human IQ means?

No, and the arithmetic shows why. An IQ is a position inside a reference sample of people. TrackingAI's number is a straight-line rescaling of a count of right answers, anchored to how Mensa Norway scores humans. On a normal curve with a standard deviation of 15, a human at 147 would sit 3.1 standard deviations above the mean, about one in 1,150 people. Whether a model that solves pattern puzzles at that level is one in 1,150 at anything else is not something the number can tell you. We make the full argument in when a model "scores 120", that number is not an IQ.

The test is also narrow. The puzzles are matrix-style pattern items of the kind we describe in Raven's matrices scores explained, which load on fluid reasoning but not on knowledge, planning or long tasks. Other leaderboards disagree with this one for reasons covered in why AI benchmark scores disagree, and for a harder, less saturated test see ARC-AGI-3. A tool such as the IQ percentile calculator will convert any number into a rank, but it cannot make a model a member of the human sample.

What should readers check before sharing an AI IQ chart?

• How many sittings stand behind the figure: an average of seven is only that when seven sittings exist
• Whether the top score is at the test's ceiling, where 35 of 35 cannot separate two models
• Whether the number is on the public test, which models may have seen, or the offline test, which has 14 scored items
• Whether the "IQ" is a rescaled count of right answers or a deviation score from a human norm group
• Whether a text-only form and a Vision form of the same model are being counted as two systems

Our IQ Metrics test reports a percentile among people, so you can see the difference between a human rank and a rescaled puzzle count on the IQ bell curve. For the wider question of machines and human intelligence, see is AI smarter than humans, and for another model-comparison index, read our guide to the most intelligent AI model.

Common questions