ChatGPT IQ: What AI Test Scores Really Mean
What is ChatGPT’s IQ—and does Claude, Grok or Gemini score higher? The evidence offers a more interesting answer than a single leaderboard: models can perform exceptionally well on some human-test tasks and poorly on others, while the meaning of those scores remains contested. This guide separates documented results from rankings that cannot be verified from the cited sources.
By MentalClarity Editorial Team · Published 17 Oct 2026 · Updated 15 Nov 2026 · 15 min read
TL;DR
- The widely reported ChatGPT score of 155 came from five verbal subtests in March 2023, not a full-scale IQ assessment. Source 2.
- A 2024 WAIS-IV study found very high working-memory and verbal-comprehension performance alongside low perceptual-reasoning performance in the models tested. Source 3.
- TrackingAI distinguishes a public Mensa Norway test from an unpublished offline test, but the cited material does not substantiate the requested July/August 2026 model scores. Source 1.
- Treat AI IQ results as test-specific benchmarks, not validated statements about a model’s human-equivalent intelligence. Source 4.
ChatGPT does not have one established, human-comparable IQ. A March 2023 report estimated a verbal IQ of 155 from five WAIS-III verbal subtests, not a complete assessment. Later research found striking differences across cognitive domains. These results describe performance under particular testing conditions; they do not establish that AI and human intelligence share a validated scale. Sources 2–4.
1. The short answer: ChatGPT IQ is not one settled number
The most defensible answer to “What is ChatGPT’s IQ?” is another question: which model, on which test, under which conditions? The sources discussed here cover an informal verbal assessment, a multidomain research study and an online testing project. These are different exercises, so their results should not be collapsed into one supposedly definitive ChatGPT IQ. Sources 1–4.
In March 2023, psychologist Eka Roivainen reported an estimated verbal IQ of 155 after presenting ChatGPT with five WAIS-III verbal subtests. That is a documented result, but its scope matters: it was not a full-scale assessment incorporating the complete verbal and non-verbal battery. Calling it simply “ChatGPT’s IQ” drops the most important qualification. Source 2.
A separate October 2024 study examined language and vision-language models using WAIS-IV tasks. Its findings were sharply uneven: very high performance on working-memory and verbal-comprehension measures coexisted with much lower perceptual-reasoning performance. That pattern gives a concrete reason to ask about abilities separately rather than treating a headline score as a complete cognitive portrait. Source 3.
TrackingAI takes another approach, comparing models on a public Mensa Norway test and an unpublished offline test. Its own description advises readers to interpret the results as performance on the selected tests and alongside the requirements of the task they want an AI to perform. That narrower framing is more useful than declaring a model universally smarter than a person. Source 1.
The exact ranges sometimes attached to these rankings need their own dated evidence. The cited TrackingAI material establishes the testing approach, but it does not contain a verifiable July/August 2026 snapshot supporting the requested scores. Consequently, this article does not present the proposed 125–135 offline range or public-test scores approaching 147 as confirmed findings.
The distinction is not an attempt to dismiss impressive results. It separates an observation—successful answers to particular questions—from a broader interpretation about intelligence. A 2026 position paper argues that applying human-calibrated psychological tests to AI requires validation before the resulting numbers can be interpreted as measurements of the same human traits. Source 4.
2. Comparing ChatGPT, Claude, Grok and other AI
A comparison table is helpful only when its numbers can be traced to a dated result. TrackingAI describes a platform that displays model rankings, question-level accuracy, answer speed, cost and changes over time. However, the source material available for this article does not establish the exact model versions or numerical entries requested for July/August 2026. Source 1.
The table therefore records an evidence gap rather than filling it with unverified figures. “Not verified” does not mean a system scored zero, performed badly or was never tested. It means this article cannot substantiate that dated entry from its cited material. There is no responsible basis here for naming an overall winner or reproducing the proposed model-version rankings.
This also applies to the requested historical claims about GPT-5.2 and o3. Without a supporting result in the permitted evidence, their proposed scores should not be repeated as facts. A publication name, an apparent date or a plausible-looking number is not a substitute for a source that actually establishes the testing conditions and result.
There is an additional scoring question worth keeping visible. Mensa Norway’s own public page describes an indicative score range of 85–145. A third-party result above that range therefore needs an explanation of the third party’s scoring method; it should not automatically be described as the standard result returned by the public website. Source 6.
For searches such as “Claude IQ”, “Grok IQ” or “Gemini IQ”, use the same verification standard. Look for an exact model version, a named test, an observation date and a clear scoring explanation before accepting a number. These are practical checks prompted by the differences between TrackingAI’s test formats and the validity concerns raised in the measurement literature. Sources 1 and 4.
Once those details are available, compare like with like. A result from the same question set and reporting method can inform a narrow benchmark comparison. It still does not establish a human-equivalent IQ or tell you which model will perform better on every other task. The distinction between benchmark performance and a validated psychological trait remains essential. Source 4.
| Model | Offline test | Mensa Norway | Notes |
|---|---|---|---|
| ChatGPT / GPT | Not verified | Not verified | No supported dated entry for the requested model version. |
| Grok | Not verified | Not verified | No supported dated entry for the requested model version. |
| Claude | Not verified | Not verified | No supported dated entry for the requested model version. |
| Gemini | Not verified | Not verified | No supported dated entry for the requested model version. |
| Perplexity | Not verified | Not verified | The requested numerical pair is not established. |
| DeepSeek | Not verified | Not verified | No supported dated entry for the requested model version. |
| Mistral | Not verified | Not verified | No supported dated entry for the requested model version. |
3. Public versus offline tests: the contamination question
TrackingAI distinguishes two testing sources. Mensa Norway is a public online test, while its offline quiz is described as having been created by a Mensa member and never published on the public internet. TrackingAI also states that verbal models receive text versions of questions, whereas vision models receive question images. Those differences belong in any interpretation of its results. Source 1.
Training-data contamination is one reason to be cautious about public benchmarks. The measurement position paper identifies contamination among the problems affecting interpretations of AI benchmark performance, alongside validity, cultural bias and sensitivity to superficial prompt changes. This supports treating prior exposure as a methodological concern; it does not establish that any particular model memorised any particular test question. Source 4.
It is useful to frame the concern as a question rather than a verdict: does success demonstrate solving a new problem, or could exposure to evaluation material have contributed? If that distinction has not been investigated, avoid claiming that a strong public-test score measures wholly novel reasoning. Equally, avoid accusing a model of memorisation simply because its result is high. Source 4.
An unpublished question set addresses one obvious route by which evaluation material could become available. Nevertheless, TrackingAI’s statement that its offline questions are absent from training data is the publisher’s assurance, not independent evidence about every model’s training history. The appropriate wording is that the project describes the questions as unpublished and outside AI training data. Source 1.
A difference between public and offline scores would not, by itself, isolate contamination as the cause. To make that conclusion, an evaluation would need to distinguish exposure from other differences between the tests and their administration. The broader concern is consistent with the position paper’s argument that benchmark numbers need a defensible measurement framework rather than an intuitive label such as intelligence. Source 4.
For readers, the practical lesson is to keep the question format visible. TrackingAI explicitly distinguishes text-based and image-based presentation, and the WAIS-IV research found a large contrast between verbal and perceptual performance. A table that mixes formats without explaining them risks hiding a difference that matters to the interpretation of model capability. Sources 1 and 3.
4. What the WAIS evidence actually shows
Roivainen’s Scientific American article is the source of a widely repeated ChatGPT IQ figure. In the March 28, 2023 account, he used five WAIS-III verbal subtests that could be presented in writing: Vocabulary, Similarities, Comprehension, Information and Arithmetic. He estimated a verbal IQ of 155 from those responses. The article is an individual testing account, not a report of a complete clinical assessment. Source 2.
Its limits are explicit. Roivainen did not administer Digit Span, the sixth verbal subtest, and did not conduct the non-verbal performance subtests. His account explains that the full-scale measure depends on the complete set of verbal and performance subtests. Therefore, “155 on an estimated verbal measure” preserves the evidence, whereas “a full-scale IQ of 155” does not. Source 2.
The October 2024 paper by Isaac Galatzer-Levy and colleagues investigated leading language and vision-language models using WAIS-IV benchmarks. The researchers focused on Verbal Comprehension, Working Memory and Perceptual Reasoning. The paper’s abstract describes the assessment as a comparison against human population norms, making it important to distinguish the numerical comparison from a claim that the same underlying human trait has been validated in AI. Sources 3 and 4.
For working memory, most models performed at or above the 99.5th percentile against the human normative comparison. The abstract characterises these strengths as storing, retrieving and manipulating tokens, including arbitrary sequences of letters and numbers. Verbal-comprehension performance was consistently at or above the 98th percentile, covering acquired information and linguistic understanding of words and their relationships. Source 3.
The perceptual-reasoning findings were very different. Multimodal models scored in the 0.1st–10th percentile range, which the authors interpreted as poor performance in interpreting and reasoning about visual information. The contrast is the central finding for an AI IQ discussion: strong results in some domains did not imply similarly strong results across the domains tested. Source 3.
These findings should remain attached to the evaluated models and the study’s date. They do not establish that every later model has the same profile, nor do they justify combining the percentile figures into an invented overall IQ. The requested additional claim of approximately 125 for text and 103 for vision is omitted because the supplied evidence does not substantiate it.
5. Progress over time needs a consistent measuring stick
The available research supports a limited statement about improvement. Within the 2024 WAIS-IV study, smaller and older model versions consistently performed worse. The authors interpreted this pattern as evidence that developments involving training data, parameter count and tuning were accompanying advances in the cognitive capabilities they assessed. That is a study-specific finding, not a universal historical IQ curve. Source 3.
TrackingAI also provides a way to explore results over time. Its description includes historical charts and comparisons of model accuracy, speed and cost. Such a platform can help readers investigate changes, but the existence of a history view does not substantiate every numerical timeline someone attributes to it. Exact historical claims still need the corresponding dated observations. Source 1.
The proposed narrative of a rise from the mid-80s in early 2024 to approximately 130 within a year and a half is not established by the cited material. Neither is the precise claim that Claude 3 was the first model to exceed 100 in March 2024. Leaving out those details preserves an important distinction between evidence of progress and an unsupported quantification of that progress.
For a meaningful timeline, first ask what stayed constant. Was the same question set used? Did the model receive images or text descriptions? Was the same scoring rule applied? These questions are especially relevant because TrackingAI uses different presentation formats and the position paper identifies prompt sensitivity and validity as challenges for interpreting benchmarks. Sources 1 and 4.
Next, separate improvement on a task from improvement in a proposed general trait. A model answering more questions correctly is a performance observation. Calling that change an increase in human-equivalent intelligence requires additional justification. The measurement critique argues that human-calibrated tests cannot simply carry their original interpretation across to non-human systems without empirical validation. Source 4.
The balanced conclusion is therefore neither “AI progress is imaginary” nor “the IQ leaderboard proves human-level general intelligence”. The 2024 results document differences associated with model age and size, while also documenting major differences between cognitive domains. Both findings deserve space in an account of progress; a single rising line would obscure the second. Source 3.
6. Why AI IQ is not directly comparable to human IQ
The fundamental issue is validity: what does the resulting number justify us saying? The 2026 position paper argues that human psychological and educational tests are theory-driven instruments calibrated to particular populations. Applying them to AI without validation risks mischaracterising what is measured. A familiar-looking score does not, on its own, establish that an AI possesses the same trait measured in humans. Source 4.
That is a more precise objection than saying machines are simply too different to test. The authors call for principled, AI-specific measurement, not an end to evaluating model performance. Their concern is the jump from benchmark success to claims about human-like intelligence or other psychological characteristics without sufficient theoretical and empirical support. Source 4.
ARC-AGI-3 offers a useful contrasting example. Its paper describes interactive, abstract, turn-based environments in which an agent must explore, infer goals, build an understanding of the environment and plan actions without explicit instructions. The benchmark focuses on fluid adaptive efficiency in novel tasks while avoiding language and external knowledge. It asks a different question from a set of written verbal subtests. Source 5.
The paper reports that humans could solve 100% of its environments, while frontier AI systems scored below 1% as of March 2026. These statements need careful wording: human solvability across the environments is not the same as saying every human participant achieved a perfect score. The AI figure belongs to the benchmark’s efficiency-based scoring framework, not an IQ scale. Source 5.
This contrast does not cancel out strong verbal results. Instead, it illustrates why several kinds of evaluation are needed before drawing broad conclusions. High performance on the WAIS-related language tasks and low performance on the reported ARC-AGI-3 evaluation concern different tasks, systems and conditions. Neither result should be stretched into a complete ranking of humans and AI across all abilities. Sources 3 and 5.
When asking whether AI is smarter than humans, replace the unrestricted comparison with a task-specific one. Which questions can it answer, which environments can it navigate and how was success measured? That approach stays closer to the observed evidence and avoids treating a human-referenced test number as a validated measure of a model’s overall intelligence. Sources 3–5.
7. How does your score compare to AI?
A reasonable starting point is curiosity about the task rather than a competition over intelligence. If you enjoy non-verbal reasoning, you can explore your performance without treating a chatbot’s reported score as a target you must reach. The methodological concern remains the same: sharing a numerical scale does not establish that two assessments, or two kinds of test-taker, support interchangeable interpretations. Source 4.
MentalClarity offers an online non-verbal reasoning test that reports a score on an IQ-style scale. It is not a clinical or diagnostic assessment, and it is not a WAIS assessment. If you would like to explore that format, the test is available at https://mentalclarity.me/iq. Its method page describes how the reported score is calculated. Source 7.
There is a broad topic overlap with visual-pattern testing, but this should not be presented as proof of direct comparability. Mensa Norway describes its own online test as 35 visual-pattern puzzles completed within a 25-minute limit. It also explicitly states that the online result is not a substitute for professional intelligence testing. Those are properties of that test, not automatically of another assessment. Source 6.
The relevant question is not merely whether two tests contain shapes or patterns. It is whether evidence supports interpreting their scores as equivalent for the comparison being proposed. The position paper makes the corresponding point about AI: a test’s original human calibration does not automatically validate applying the same trait interpretation to an artificial system. Source 4.
For that reason, this article does not invite you to subtract your score from an AI leaderboard entry and interpret the difference as an intelligence gap. Nor does it label a higher result as proof that you are smarter than a named model. The cited evidence does not establish a common validated measurement scale for that conclusion. Source 4.
A more grounded comparison is descriptive. Ask whether the tasks were verbal or visual, whether the same materials were used and whether administration differed. TrackingAI’s distinction between text versions and question images shows why these details belong beside a score. Understanding the task makes the comparison more informative than treating the largest displayed number as the whole story. Source 1.
8. How to read the next AI IQ headline
Start by identifying the claim’s level. “The model answered these items correctly” is narrower than “the model scored at this human-reference percentile”, which is narrower than “the model has this human-equivalent IQ”. The measurement critique challenges the move from observed benchmark performance to human-like psychological traits without validation. Keeping those steps separate makes a headline much easier to assess. Source 4.
Then check whether the number belongs to a complete assessment or a subset. The 2023 ChatGPT example shows why this matters: five written verbal subtests supported an estimated verbal result, not the full-scale conclusion often implied by the shortened headline. Preserve the words that describe the test’s scope even when they make the summary less dramatic. Source 2.
Check the date and the evaluated system next. The WAIS-IV paper reports a particular model sample in 2024; the ARC-AGI-3 paper identifies its frontier-model result as of March 2026. Neither should be silently converted into a claim about every currently available chatbot. A result is most informative when its date and testing context remain attached. Sources 3 and 5.
Finally, ask whether the evidence supports the decision you want to make. TrackingAI advises considering its selected-test results alongside the requirements of the task you want an AI to perform. That is a useful boundary: an IQ-style ranking can be one observation, but it should not substitute for evaluation against your actual requirements. Source 1.
- Ask for the exact test and the scope of the result: full assessment, selected subtests or an online puzzle set. Sources 2, 3 and 6.
- Check how questions were presented, especially whether models received text or images. Source 1.
- Treat contamination as a concern to investigate, not a proven explanation for every high score or score difference. Source 4.
- Keep benchmark performance separate from claims about human-equivalent intelligence. Source 4.
Frequently asked questions
Is ChatGPT smarter than humans?
There is no single supported answer across all abilities. A 2024 WAIS-IV study found exceptionally high working-memory and verbal-comprehension performance in the models tested, alongside low perceptual-reasoning performance. Separately, ARC-AGI-3 reported a large human–AI performance gap on its novel interactive environments as of March 2026. These are task-specific findings, not a universal intelligence ranking. Sources 3 and 5.
Which AI has the highest IQ?
The cited evidence does not establish one overall winner. TrackingAI compares models on public Mensa Norway and unpublished offline questions, but the material available here does not substantiate the requested July/August 2026 numerical rankings. Even a verified leader would lead on a particular test and reporting method, not necessarily possess the highest human-equivalent intelligence. Sources 1 and 4.
Can an AI join Mensa?
A chatbot’s reported score on an online Mensa-branded test does not establish membership eligibility. Mensa Norway explicitly says its public online test is not a substitute for professional intelligence tests, including those administered by Mensa. The sources cited here do not establish a membership policy for AI, so they do not support claiming that a model qualifies to join. Source 6.
What IQ does Claude have?
This article cannot substantiate one Claude IQ or the requested dated leaderboard score. The available TrackingAI material describes its testing methods but does not establish that numerical entry. A useful answer would need an exact model version, test and date; even then, it would describe test performance rather than a validated human-equivalent IQ. Sources 1 and 4.
What IQ does Grok have?
The requested Grok figures are not verified by the cited material, so no numerical score is assigned here. Any reported result should identify the model version, observation date and whether it came from public or unpublished questions. If a third-party Mensa Norway score exceeds the public website’s stated 85–145 range, its scoring method also needs explanation. Sources 1 and 6.
Sources
- AI IQ Scores & Model Rankings | TrackingAI — trackingai.org
- I Gave ChatGPT an IQ Test. Here's What I Discovered | Scientific American — scientificamerican.com
- The Cognitive Capabilities of Generative AI: A Comparative Analysis with Human Benchmarks — arxiv.org
- Position: Stop Evaluating AI with Human Tests, Develop Principled, AI-specific Tests instead — arxiv.org
- ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence — arxiv.org
- IQ Test Made by Mensa Norway - Mensa Norway — test.mensa.no
- How the Score Is Calculated — MentalClarity IQ Method — MentalClarity
Editorial note: this article is general information, not medical or psychological advice. Sources were checked before publication and are re-checked weekly.