Are You Smarter Than AI?

weirdly accurate.

Are You Smarter Than AI?

Are you smarter than AI at ARC-AGI puzzles?

Not any more, on these puzzles, though people held the lead for five years. When François Chollet published ARC in 2019, its puzzles were easy for people and very hard for machines. In the H-ARC study (2024), people solved about 60% of the evaluation puzzles within two tries, while GPT-4o, in ARC Prize’s run, solved 11.4%. Today’s top models solve about 95–98% of the same 400 public puzzles.

This ARC-AGI test gives you six puzzles from that public evaluation set, ordered from easier to harder by how people did on them. Each one shows two or three examples of a grid turning into another grid; you work out the rule and paint the answer on a peg board, choosing colours, filling areas and resizing the grid. As in ARC Prize testing you get two tries and are told only right or wrong; after a second miss the answer is shown. Your score is the number of puzzles you solve. It is placed among people with a Rasch model fitted to 946 H-ARC volunteers who did puzzles from the same set, and set beside which of the six GPT-4o and three 2026 models solved.

It is not an IQ test. Six puzzles is a tiny sample, and one lucky guess moves your place a long way. Nobody in H-ARC did these exact six (each person tried about five random puzzles), so the percentile comes from the model, not from a head count. The model scores are flattering in one way: these puzzles have been online since 2019, so newer models may have seen them in training. ARC Prize’s official scores use puzzles kept off the internet.

What the ARC-AGI test measures

One thing: how well you learn a new rule from two or three examples, which is what ARC was built to measure.

The task

  • Six puzzles from the ARC-AGI-1 public evaluation set (Abstraction and Reasoning Corpus, François Chollet, 2019), each with a single test grid of at most 10 × 10
  • Ordered by difficulty estimated from the H-ARC human data, with a stop after three
  • Paint the answer with ARC’s ten colours: tap or drag across cells, fill a connected area, copy the input, clear, and drag the corner to set the size
  • Two tries per puzzle, told only right or wrong; handing in the same wrong grid twice does not use up a try

Your result

  • How many of the six you solved, and on which try
  • How many people out of 100 you would beat, from a Rasch model fitted to H-ARC’s evaluation-set participants
  • For each puzzle, how many of the H-ARC volunteers who tried it solved it within two tries
  • Which of the six GPT-4o and three 2026 models (Gemini 3.1 Pro, GPT-5.4 Pro, Claude Opus 4.8) solved, and their scores on all 400 public puzzles
  • How ARC changed after reasoning models: o3-preview in 2024, ARC-AGI-2 and ARC-AGI-3
  • A shareable card and a link that challenges a friend to solve more

Before you start

  • About five minutes, with a stop halfway; free, with no sign-up
  • Works on a phone, where you switch between the examples and your board; a wide screen shows both at once
  • No sound needed
  • No reading needed either: every puzzle is pictures of coloured squares

ARC-AGI test FAQ

Is this test free? Do I need to sign up to see my result?

Yes, it’s free, and there’s nothing to sign up for: no account, no email, no payment. Your result appears on this page the moment you finish, and you can share it or play again straight away.

Can humans still beat AI at ARC-AGI?

On the original ARC-AGI-1 public puzzles, mostly not: in ARC Prize’s 2026 runs, Gemini 3.1 Pro, GPT-5.4 Pro and Claude Opus 4.8 solved 95–98% of them. On the harder ARC-AGI-2, which keeps only puzzles that at least two of 400 people solved within two tries, the best verified commercial model scored 37.6% in December 2025.

How many ARC puzzles does a typical person solve?

About six in ten. In the H-ARC study by LeGris, Vong, Lake and Gureckis, volunteers solved about 60% of the evaluation puzzles they tried within two tries, and 48% at the first try. Of all 800 public ARC puzzles, 790 were solved by at least one person within three tries.

What is a good score on this ARC-AGI test?

Four of the six is about the middle. The Rasch model fitted to H-ARC predicts that about half of people would solve four or five of these six puzzles, and about 11% all six. The puzzles are ordered from easier to harder, so the last ones decide most results.

What is ARC-AGI?

ARC-AGI is a set of 800 public puzzles, the Abstraction and Reasoning Corpus, that François Chollet published in 2019 with his paper On the Measure of Intelligence. Each puzzle has its own rule, shown in two or three examples and built only from things people know without being taught, such as objects, counting and symmetry. Chollet’s aim was to measure how efficiently a mind learns a new skill, not how much skill it already has.

Is this ARC-AGI test an IQ test?

No. Six puzzles is far too few to measure general intelligence, and the ARC-AGI test gives no IQ score. Its percentile is a model estimate from H-ARC volunteers, who were adults aged 18 to 77 recruited online, so it compares you with that group on these six puzzles only.

How did AI get so good at ARC?

Reasoning models changed it. In December 2024 OpenAI’s o3-preview scored 75.7% on ARC Prize’s hidden puzzles at about $26 of computing per puzzle, and 87.5% with 172 times more computing, about $4,560 a puzzle. ARC Prize put the cost of paying a person at roughly $5 per puzzle.

Where the numbers come from

Human rates are our recount of H-ARC’s public summary data, counting two tries as ARC Prize does (H-ARC itself allowed three). The percentile comes from a Rasch model we fitted to the evaluation-set participants; each puzzle’s own count comes from only 8–12 people, so it is rough. Model scores are our recount of ARC Prize’s published answer files on the 400 public evaluation puzzles. The site does not publish its own players’ data.

How the score is worked out

Two tries

A puzzle counts as solved if your first or second answer matches the test output exactly, cell for cell and size included. This is how ARC Prize scores models, and how the human rates here are counted.

Rasch model

A statistical model that gives each person one puzzle-solving ability and each puzzle one difficulty. Fitted to the two-try results of 946 H-ARC volunteers on the evaluation set, it predicts how many of these six a person would solve, and your percentile is the share of people predicted to solve fewer than you.

Model results

Recounted by us from the answer files ARC Prize publishes for each model on the 400 public evaluation puzzles, with two tries per puzzle; a puzzle with no answer file counts as unsolved.