Short answer: ChatGPT has no IQ in the human sense, but its models post high scores on IQ-style puzzles. Maxim Lott's TrackingAI recorded ChatGPT Pro at 148 on the public Mensa Norway test in October 2025; on September 28, 2026, 36Kr reported OpenAI's GPT-6 Astra at 151, that test's ceiling. On Lott's private offline test, top scores run near 130 to 135. These are puzzle results, not human-comparable IQs.
At a glance
| Best public-test result (OpenAI) | 151 on Mensa Norway, GPT-6 Astra, reported September 28, 2026 (36Kr, citing TrackingAI) |
|---|---|
| Best offline-test result | 135, GPT-5.6 Terra with image input, same report |
| First widely cited score | Verbal IQ 155 on five WAIS-III subtests (Roivainen, Scientific American, 2023) |
| Tracker | TrackingAI, run by journalist Maxim Lott |
| Measured rate of gain | About 2.5 IQ points a month, May 2024 to October 2025 (Lott) |
| Comparable to a human IQ? | No: different conditions, contamination risk, no AI norming sample |
The Scores, Date by Date
AI scores change by the month, so every figure here carries the date it was reported and the person who reported it. The longest-running record belongs to Maxim Lott, a journalist who has tested chatbots on IQ-style puzzles since at least early 2024 and publishes results on his site TrackingAI and his newsletter, Maximum Truth.
| Date reported | Model | Test | Result | Source |
|---|---|---|---|---|
| 2023 (Scientific American) | ChatGPT | Five WAIS-III verbal subtests, typed | Verbal IQ 155 | Eka Roivainen, Oulu University Hospital |
| 2023 (Nature Human Behaviour) | GPT-3 (text-davinci-003) | Text version of Raven-style matrix problems | Matched or exceeded human participants in most settings | Webb, Holyoak and Lu |
| September 14, 2024 | OpenAI o1-preview | Mensa Norway | 25 of 35 correct; an IQ "as high as 120" | Maxim Lott, Maximum Truth |
| April 19, 2025 | OpenAI o3 | Mensa Norway / offline test | 136 / 116 | Maxim Lott, Maximum Truth |
| October 31, 2025 | ChatGPT Pro | Mensa Norway / offline test (text) | 148, on 34 of 35, twice / 117 | Maxim Lott, Maximum Truth |
| September 28, 2026 | GPT-6 Astra; Claude Fable 5.1 | Mensa Norway | 151, all 35 correct, the top of the scale | 36Kr, citing Lott's TrackingAI |
| September 28, 2026 | GPT-5.6 Terra (image input) | Offline test | 135, against a maximum of about 136 | 36Kr, citing Lott's TrackingAI |
Read the table as a trend, not a leaderboard. Model names change, companies adjust settings after launch, and a single week's result can swing by several points. Progress is not smooth: Lott reported that AI scores barely moved for about six months before o1-preview arrived in September 2024.
How TrackingAI Tests a Chatbot
Lott uses two instruments. The first is the Mensa Norway test, a public online quiz of 35 visual-pattern items with a 25-minute limit; our guide to the Mensa Norway test explains its format. The second is an offline test written from scratch by a Mensa member identified as Jurij and never posted on the internet.
Lott normed the offline items against the Mensa Norway items by having his readers take both, so that a 100 means roughly the same on each. The September 2024 calibration rested on about 40 readers. Lott reported that the Mensa Norway test itself was normed on 80 people with official IQ scores, a claim from its creator that he did not verify. According to 36Kr's September 2026 report, TrackingAI tests more than 30 models a week and averages each model's last seven attempts. Some models see the puzzles as images; others receive them described in words, which has so far produced higher scores.
The Contamination Problem
A language model learns from a vast scrape of public text. If a test's questions and answers sit somewhere in that scrape, a high score may reflect memory rather than reasoning. Lott called this the most common objection to his early results, and he notes that Mensa Norway answers exist online.
The gap between his two tests is the evidence. In October 2025, ChatGPT Pro scored 148 on the public test and 117 on the private one. Lott wrote that contamination partly explains why top systems near the public test's ceiling while scoring about 130 offline.
Researchers have documented the problem on other benchmarks. Shahriar Golchin and Mihai Surdeanu, in a 2023 preprint, built a method to detect test items a model has memorized and concluded that GPT-4 had seen parts of three standard datasets, including AG News and XSum. Any test with published answers carries the same risk, which is why the offline score is the more meaningful number.
Why an AI Score Is Not a Human IQ
A human IQ is a deviation score. It compares one person with a large, representative sample of people the same age, scaled to a mean of 100 and a standard deviation of 15. No such sample exists for machines, and the human sample was never designed to rank them.
- Uneven profiles. Roivainen gave ChatGPT the Vocabulary, Similarities, Comprehension, Information and Arithmetic subtests of the WAIS-III. It earned a Verbal IQ of 155, yet he observed that it failed a simple riddle. He skipped Digit Span, a working-memory task, because the chatbot lacks that memory system, and nonverbal subtests were not possible.
- Different conditions. A model does not tire, can retry, and runs with a compute budget the vendor sets. Time limits built for human attention do not constrain it in the same way.
- Ceiling effects. 36Kr noted that Mensa Norway reports human scores only up to 145. In December 2025 Lott said he planned to improve his test's granularity above 130, where it separates top systems poorly.
- Breadth. Lott quoted the economist Tyler Cowen's view that such tests underweight a model's extreme breadth of knowledge, the thing it does best.
The adult intelligence battery behind most clinical scores is described in WAIS-IV explained.
What the Scores Do Show
The trend is real even if the number is not an IQ. Lott measured an average gain of about 2.5 IQ points a month for the leading AI between May 2024 and October 2025, from the mid-80s to about 130. He projected in October 2025 that leading systems would top out on human IQ tests in late 2027.
Academic work points the same way. In 2023, Taylor Webb, Keith Holyoak and Hongjing Lu reported in Nature Human Behaviour that GPT-3 matched or surpassed human participants on most analogy tasks, including a text version of matrix problems modeled on Raven's Progressive Matrices. They called the ability emergent: it arose without training on those specific tasks.
What the scores cannot show is whether a system understands, plans or judges as a person does. Alan Turing posed that question in 1950 with a conversation test, not a puzzle sheet; see what is the Turing test and the history of artificial intelligence.
How the Numbers Map Onto Human Scores
On a human scale, an offline score of 130 would place a person at about the 98th percentile, the traditional Mensa threshold. See what an IQ of 130 means. A 151 would be rarer than 1 in 1,000. Those comparisons describe the scale, not the machine: the model reached its score under conditions no human test taker faces.
For human scores, the same caution about format applies in reverse. Quick online quizzes are useful practice and poor measurement; read are online IQ tests accurate. Our free IQ test is a practice screener, not a clinical assessment. Every related guide is collected in IQ guides.
Frequently asked questions
What is ChatGPT's IQ?
It has no IQ in the human sense. On the public Mensa Norway test, Maxim Lott recorded ChatGPT Pro at 148 in October 2025, and 36Kr reported on September 28, 2026, that OpenAI's GPT-6 Astra had reached the test's ceiling of 151. On Lott's private offline test, the top OpenAI result in that report was 135.
Is ChatGPT smarter than humans?
On narrow puzzle tests, current models outscore most people. That does not make them smarter in general. They can fail simple riddles, depend on settings their makers control, and may have seen published test answers during training. An IQ-style score measures one kind of task, not overall intelligence.
Which AI has the highest IQ score?
Rankings change monthly. In a September 28, 2026 report citing TrackingAI, 36Kr said two models, OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1, scored 151 on Mensa Norway, while GPT-5.6 Terra with image input led the offline test at 135. Check the tracker for the latest week.
Why do AI models score lower on the offline test?
Because its questions have never been published, so models are unlikely to have memorized them. The public Mensa Norway test has answers online, and a model may have absorbed them in training. Lott attributes much of the gap between the two scores to that exposure.
Can ChatGPT take a real clinical IQ test?
Only in part. A psychologist can type verbal subtests to it, as Eka Roivainen did with the WAIS-III, but working-memory and hands-on performance subtests do not translate. The result is a partial profile scored against human norms, not a full-scale IQ.
Sources
- Maxim Lott, Massive breakthrough in AI intelligence: OpenAI passes IQ 120 (Sept. 14, 2024) — o1-preview 25 of 35 on Mensa Norway; offline test construction and 40-reader norming; contamination objection; six-month plateau
- Maxim Lott, Skyrocketing AI Intelligence: ChatGPT's o3 sets new record (Apr. 19, 2025) — o3 136 Mensa Norway, 116 offline; Tyler Cowen's breadth caveat
- Maxim Lott, AI progress continues, as IQ scores rise linearly (Oct. 31, 2025) — ChatGPT Pro 148 Mensa Norway, 117 offline; 2.5 points a month; contamination gap; late-2027 projection
- Maxim Lott, Charting AI's Rise: 2025's Intelligence Breakthroughs (Dec. 31, 2025) — Leading US model at 130 on the text version of the test; planned finer scoring above 130
- 36Kr, AI Outperforms 99.97% of Humans: Scores 151 on Mensa IQ Test (Sept. 28, 2026) — GPT-6 Astra and Claude Fable 5.1 at 151; GPT-5.6 Terra 135 offline; 35 items, 25 minutes; seven-attempt average
- TrackingAI IQ tracker — Live leaderboard and method notes; the 36Kr model scores could not be read from the page text
- Eka Roivainen, I Gave ChatGPT an IQ Test. Here's What I Discovered (Scientific American, 2023) — Verbal IQ 155 on five WAIS-III subtests; riddle failures
- Webb, Holyoak and Lu (2023), Emergent analogical reasoning in large language models, Nature Human Behaviour 7(9) — GPT-3 matched or exceeded humans on matrix-style analogy problems
- Golchin and Surdeanu (2023), Time Travel in LLMs: Tracing Data Contamination, arXiv — Evidence that GPT-4 had memorized parts of standard test datasets
Editorial standard: IQ figures for historical figures are retrospective estimates and are labeled as such; scores for living people appear only when the person disclosed them or a reputable outlet reported them. Corrections: contact the editors.
