How Smart Is AI? Five Ways It Is Measured, and What Each One Says

How smart is AI? There is no single answer, because five common yardsticks report five different kinds of number. Read on 12 November 2026, one yardstick reports tasks that take a human expert up to 17.4 hours, another reports 96.3% on graduate-level science questions, and a third shows models getting about 30 of 35 puzzles right on a Mensa Norway test. On new games built for people, frontier AI scored under 1% in March 2026. None of these numbers is an IQ.

Below: what each yardstick counts, its latest figure with the date we read it, why the numbers disagree, and why none of them converts into IQ. This page explains how AI is measured. It does not give a verdict. For that, see our answer to is AI smarter than humans.

How smart is AI right now, and why is there no single number?

AI looks very strong on tests with one correct answer and, in March 2026, looked weak on games built to be new. People ask how smart is AI compared to humans, and the honest reply is: on which test? Five yardsticks are in common use. Each reports a different unit, so you cannot add them up, rank them against each other or turn any of them into an IQ.

• Task length (METR): the hours of human expert work in tasks an AI completes half the time. Up to 17.4 hours on 12 November 2026, against 4.0 minutes for GPT-4 in March 2023
• Expert exams (GPQA Diamond and Humanity's Last Exam): percent correct. On 12 November 2026, 96.3% on GPQA Diamond and 74.3% on Humanity's Last Exam, both as listed by Artificial Analysis
• New games (ARC-AGI-3): the share of game environments solved, scored against human action counts. In March 2026, humans 100% and frontier AI under 1%
• IQ-style puzzles (Mensa Norway test, run by TrackingAI): puzzles right out of 35. The average sitting scored 12.0 in May 2024 and 29.6 in October 2026
• Conversation (the Turing test): the share of five-minute chats in which the AI is judged to be the human. 73% for GPT-4.5 with a persona prompt, in a PNAS paper of May 2026

How long a task can AI finish?

At most about 17.4 hours of expert work, but METR calls readings above 16 hours unreliable. METR's 50% time horizon is the length of task, measured by how long a skilled human takes, that an AI agent completes about half the time. It measures how hard a task is, not how long the AI works. METR says agents are typically several times faster than people on tasks they complete.

In METR's data file for its current task suite, read on 12 November 2026, GPT-4 (released 14 March 2023) sits at 4.0 minutes. The top entry is an early version of Claude Mythos Preview, released 7 April 2026, at 1,044.8 minutes, which is 17.4 hours, with an interval of 8.5 to 55.1 hours. METR's page says measurements above 16 hours are unreliable with its current tasks. The page was last updated on 8 September 2026 and is no longer actively updated.

METR's paper (arXiv 2503.14499, March 2025, revised July 2026) found a 50% time horizon of around 50 minutes for Claude 3.7 Sonnet and a doubling about every seven months since 2019. The data file gives a doubling time of 128.7 days since 2023, with an interval of 104.4 to 158.0 days. The tasks are mostly software engineering, machine learning and cybersecurity, so the figure does not describe office work in general. Our guide to the METR time horizon goes through the numbers.

What do expert exams like GPQA Diamond show?

They show how often a model picks the right answer on questions written by PhD-level experts, and the top models now score above 95% on GPQA Diamond. The Diamond set has 198 four-option multiple-choice questions in biology, physics and chemistry, so random guessing scores 25% (Epoch AI). On Artificial Analysis on 12 November 2026, GPT-6 Astra (Xhigh) led at 96.3%, then GPT-6 Astra (Max) at 96.1% and Gemini 3.8 Flash (High) at 95.3%.

The human side is older and smaller. In the 2023 GPQA paper (Rein and colleagues, arXiv 2311.12022, 20 November 2023), PhD experts reached 65% on the 448-question set, or 74% after discounting mistakes they spotted afterwards. Skilled non-experts with web access reached 34%, and the best GPT-4 baseline reached 39%. Epoch AI adds that PhD-level experts recruited by OpenAI scored 69.7% on Diamond. With 198 questions, one question is about 0.5 points, so the top three differ by about two questions. Our GPQA Diamond guide covers the detail.

Humanity's Last Exam (HLE) was built to be harder. It has 2,500 closed-ended questions from nearly 1,000 experts at over 500 institutions in 50 countries, and a question got in only if frontier models failed it (Scale AI). On 12 November 2026, Artificial Analysis showed Claude Fable 5.2 (Max) first at 74.3%, then Claude Fable 5.2 (Xhigh) at 68.9%. It files HLE under its "Legacy" evaluations, and other boards give other top scores, as our Humanity's Last Exam leaderboard guide shows. Neither exam has a norm group of people behind it, so neither score is an IQ.

What does a puzzle yardstick like ARC-AGI-3 show that an exam does not?

It shows how well an AI learns a new task from scratch, which an exam of known facts cannot. ARC-AGI-3 drops an agent into a small turn-based game with no instructions, and its scoring compares the agent's number of actions with human baselines. The technical report (arXiv 2603.24621, 24 March 2026) says humans can solve 100% of the environments, while frontier AI systems "as of March 2026, score below 1%".

ARC Prize's page, read on 12 November 2026, says a 100% score means AI agents can beat every game as efficiently as humans. We could not read the current official leaderboard, because it loads by script, so we state no official score after March 2026. Our report on ARC-AGI-3 explains what the benchmark does and does not measure.

Why does one model score 100% on one test and near 0% on another?

Because the tests count different things, and some were built against current models or can be tuned to. A September 2026 preprint, Kepler (arXiv 2610.00834), reports a server-verified 100.00 on all 25 public ARC-AGI-3 games for one frozen Claude Opus 5 setup. The setup uses its own harness, which is the software wrapped around the model to choose its moves. The authors conclude that the public-set score alone has limited discriminative value.

That is one team's preprint, not a peer-reviewed result, and it covers the public games, not the hidden ones. HLE works the other way: its questions were picked because frontier models failed them, so its scores started low by design. A high score on one test says little about the next. Our article on why AI benchmark scores disagree shows the same model scoring differently under different grading rules.

How does an IQ-style test score AI?

It counts correct puzzles and then converts the count with a formula, and a formula is not a norm. TrackingAI runs AI models on the public Mensa Norway test, which has 35 visual-pattern puzzles and a 25-minute limit for people. TrackingAI turns the count into an "IQ" with a straight line: IQ = 63.5 + 3 x (correct – 5.833). Each extra correct answer adds 3 points, 18 of 35 equals 100 and 35 of 35 equals 151 (its chart code, read 12 November 2026). Mensa Norway's own scale ends at 145, and it calls its result an indication, not a substitute for professional tests.

TrackingAI's public log holds 1,931 Mensa Norway sittings from 9 May 2024 to 9 November 2026. The average sitting scored 12.0 of 35 in May 2024 (94 sittings) and 29.6 in October 2026 (32 sittings). On TrackingAI's line, that is about 82 and about 135. Only 13 sittings, 0.67%, were a perfect 35. Our report on AI scores on the Mensa Norway test looks at the 151 ceiling.

Do chat tests like the Turing test measure intelligence?

They measure how human a conversation feels. In a PNAS paper published on 19 May 2026, GPT-4.5 prompted to act like a person was judged to be the human in 73% of five-minute chats, and LLaMa-3.1 in 56%. Without those prompts, the same models fell to 38% and 36%, and ELIZA, a chatbot from the 1960s, scored 23%. The authors say interrogators focused on style and social-emotional cues more than on traditional notions of intelligence. We cover the study in has AI passed the Turing test.

Is any of this an IQ?

No. An IQ places a person inside a reference group with a stated average and spread, and on most tests the average is 100 and one standard deviation is 15. Our IQ percentile calculator shows what a score means against that group. The five yardsticks above report hours, percent correct, a share of games, a puzzle count and a win rate, and none has a norm group of people behind it. Any "AI IQ" you see is a conversion, like TrackingAI's line.

The same goes for a product. If you ask how smart is ChatGPT, the answer depends on the model and the test, and our page on the ChatGPT IQ score covers that case. For size rather than test scores, see our research on how big AI is compared to a brain.

To measure your own reasoning against other adults, take the IQ Metrics test.

Common questions