AI IQ: What IQ Tests Say About ChatGPT and Other Models

What is AI's IQ? Published IQ-test scores for ChatGPT, o3 and other models, public vs offline tests, and why IQ is a weak measure of AI.

By Izhar Ahmad Danish · · 3 min read

AI IQ: What IQ Tests Say About ChatGPT and Other Models: Time of Superintelligence cover image
On this page
  1. What IQ do AI models score?
  2. Why do the two tests disagree?
  3. What is ChatGPT’s IQ?
  4. Why is IQ a weak measure of AI?
  5. Is ChatGPT smarter than Einstein?
  6. Our take
  7. Frequently asked questions
  8. Sources

Search for AI IQ and you’ll find headlines claiming ChatGPT has an IQ of 130, 140 or more. Some of those numbers are real measurements; many are guesses. This page lists the IQ-test results we could verify, explains why the same model can score very differently, and argues that IQ is a weak way to measure AI.

Short answer: AI models now score above average on human IQ tests. In April 2025, OpenAI’s o3 scored 136 on the public Mensa Norway test, higher than about 98% of people. But on an offline test never posted online it scored 116. IQ is built for humans, so these scores say little about how well AI does real work.

What IQ do AI models score?

Short answer: from below average to gifted, depending on the model and the test.

The most-cited measurements come from TrackingAI.org, which gives models two tests: the public Mensa Norway online test and an Offline Test written by a Mensa member and never published, so it can’t be in training data.

Model Mensa Norway (public) Offline test (unpublished) Gap
OpenAI o3 136 116 20
OpenAI o1 Pro (text) 122 107 15
Meta Llama 4 Maverick 106 97 9
OpenAI GPT-4o 95 64 31
Most Apache-licensed open models roughly 60–90 — —

Data: TrackingAI.org, as reported by CryptoSlate on 17 April 2025. Human average = 100. Newer models have since been added to TrackingAI’s live charts; check the site for current numbers.

Why do the two tests disagree?

Short answer: contamination. Public questions may be in the model’s training data.

Every model in the table scores lower on the unpublished test, by 9 to 31 points. The likeliest reason is that the Mensa Norway questions, or discussions of their answers, are on the open web. The offline score is the better guide to fresh reasoning. Input format matters too: o1 Pro scored 122 on the text version but 86 when given the question images.

What is ChatGPT’s IQ?

Short answer: there’s no single number. “ChatGPT” is a product that has run many models.

GPT-4o, the default ChatGPT model for much of 2024–25, scored 95 on Mensa Norway, roughly average. Reasoning models such as o3 scored far higher. Any article quoting “ChatGPT’s IQ” without naming the model and test is not telling you much.

Why is IQ a weak measure of AI?

Short answer: IQ was designed to predict human abilities, and those correlations don’t carry over to AI.

  1. It’s calibrated on humans. An IQ of 130 means “better than about 98% of people.” That ranking only predicts other abilities because, in humans, skills tend to rise together. In AI they don’t: Stanford’s 2026 AI Index found the top model read analog clocks correctly only 50.6% of the time, against about 90% for people.
  2. It can be memorised. Public tests leak into training data, as the score gap above shows.
  3. It has a ceiling. IQ scales stop being meaningful far above the human range, so they can’t measure superintelligence.
  4. It misses what matters. The bigger question is whether AI can finish real tasks reliably. On that, AI still trails: agents fail about one in three structured tasks.

Better signals include ARC-AGI (novel puzzles designed around what humans find easy), Humanity’s Last Exam, and agent benchmarks. We compare them in is AI smarter than humans?

Is ChatGPT smarter than Einstein?

Short answer: no. A puzzle score isn’t scientific genius.

Einstein’s importance was producing ideas nobody had, not his test score. AI can now beat most people at matrix puzzles and match olympiad medallists at maths, but it hasn’t independently produced a breakthrough on that scale. If that happens, it will be a sign we’re close to AGI.

Our take

AI IQ scores are a fun headline and a poor yardstick. Use the offline-test numbers if you want a rough comparison, and ignore claims that don’t name the model and the test. To judge how capable AI really is, look at whether it can do whole tasks without supervision; that’s where the meaningful progress, and the remaining gap, is.

Enjoyed this? Get the SI Briefing

One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.

Frequently asked questions

What is ChatGPT's IQ?

It depends on the model and the test. In April 2025 TrackingAI measured OpenAI's o3 at 136 on the public Mensa Norway test but 116 on an offline test; GPT-4o scored 95 and 64. Newer models score higher, but treat any single number with caution.

What IQ is superintelligence?

There is no meaningful number. IQ is scaled to the human population, with a mean of 100 and a standard deviation of 15, so scores far beyond the human range can't be measured. Superintelligence would exceed humans across the board, which a single IQ score can't capture.

Is ChatGPT smarter than Einstein?

No. A chatbot can outscore most people on puzzle tests and recall far more facts, but Einstein produced new physics. No AI has yet done original scientific work at that level on its own.

Is 97 IQ dumb?

No. IQ tests are scaled so 100 is average and about two-thirds of people score between 85 and 115. A score of 97 is squarely average.

Can you give AI an IQ test?

Yes, and people do, usually with text or image versions of matrix-reasoning tests. The results are interesting but limited: AI may have seen the questions in training, and IQ doesn't predict AI's real-world reliability the way it does for humans.

4 Sources

  1. TrackingAI.org: AI IQ test methodology (Mensa Norway and Offline Test)
  2. CryptoSlate: OpenAI's o3 scores 136 on Mensa Norway test (17 Apr 2025)
  3. Stanford HAI, 2026 AI Index: Technical performance
  4. ARC Prize, ARC-AGI results
Before you go

Super intelligence is moving fast.

Don’t miss what happens next

Get one short email a week with what changed and why it matters. No spam. Unsubscribe anytime.