Are You Smarter Than AI?

The best AI models now score 133 on a private IQ test they have never seen, higher than about 98.6% of people. Enter your IQ to see which of 17 models you beat.

AI scores updated 1 October 2026 · Source: TrackingAI.org

Compare on

On the private test AI has never seen

You beat 4 of 17 AI models

An IQ of 120 is higher than about 91% of people. The top model, GPT 6 Astra Ultra, scores 133 on the private test.

  1. GPT 6 Astra Ultra

    OpenAI · vision

    133

  2. Claude-5.5 Opus MAX

    Anthropic · vision

    133

  3. GPT 5.6 Terra Ultra

    OpenAI

    132

  4. Muse Glimmer

    Meta

    123

  5. GLM 5.2

    Z.ai

    123

  6. You

    Rank 14 of 18

    120

  7. Claude-5 Sonnet

    Anthropic

    115

  8. Bing Copilot

    Microsoft

    87

The private test was written by a Mensa member and has never been online, so no AI could have memorised it. Its scale tops out at about 136. Faded rows are models your score beats.

Already know your score from another test? See its percentile.

AI IQ leaderboard

Every current model, ranked by the private test. Scores as of 1 October 2026.

ModelMakerPrivate testPublic testLast tested
GPT 6 Astra UltravisionOpenAI13314830 September
Claude-5.5 Opus MAXvisionAnthropic13314428 September
GPT 5.6 Terra UltraOpenAI13214330 September
GPT 6.1 Sol UltravisionOpenAI13113629 September
Claude-5.1 FablevisionAnthropic1301453 September
GPT 6 Luna MaxvisionOpenAI1301391 October
Kimi K3Moonshot AI130–29 September
Grok 4.7 HighxAI12914224 September
Gemini 3.8 FlashGoogle1291421 September
Claude-5.5 OpusAnthropic1281441 October
Qwen 3.8 MaxAlibaba12614310 September
Muse GlimmerMeta12313018 September
GLM 5.2Z.ai123–10 September
Claude-5 SonnetAnthropic11513630 September
Bing CopilotMicrosoft8710110 September
DeepSeek V4 ProDeepSeek8410310 September
ManusButterfly Effect8311810 September

How fast AI got smarter

The best single score on each test, every quarter since TrackingAI began testing in 2024.

6080100120140160Average person2024 Q22024 Q42025 Q22025 Q42026 Q22026 Q4
Public test, best runPrivate test, best run

Why there are two scores

The public test is Mensa Norway's online test: 35 visual matrix puzzles that anyone can take. Because it is online, AI models may have met the puzzles during training. Its scale tops out at 151, and the best models are now close to it.

The private test was written by a Mensa member and has never been published, so no model could have memorised it. Most models score 10 to 20 points lower on it, which is the fairer measure of how well they reason about new problems. Its scale tops out at about 136.

That is why we rank by the private test. If your IQ is above a model's private score, you solved fresh matrix puzzles better than it did.

What AI IQ scores do and don't mean

Matrix puzzles are a fair test of pattern reasoning, and the progress is real. In early 2024 the best model scored around 100. Two years later the best models beat almost everyone. But keep four things in mind:

  • No time limit. People take these tests against the clock. Models do not get tired, nervous or rushed.
  • Words versus pictures. Text-only models get each puzzle described in words, which is a different task from seeing it.
  • A simple conversion. Each correct answer is worth about 3 points, and random guessing gives about 63.5. That is a useful mapping, not a test normed on thousands of people.
  • One narrow skill. A high score on matrix puzzles does not mean a model understands the world as you do. They still make mistakes no person would.

Other projects estimate "AI IQ" by mapping coding and maths benchmarks onto an IQ scale. Those numbers use a different method and are not comparable with the scores here.

How the scores are measured

The scores come from TrackingAI.org, run by Maxim Lott, which gives every major model the same two tests each week. We use TrackingAI's own conversions:

  • Public test: IQ = 63.5 + 3 × (raw score − 5.833), 35 items; maximum 151.
  • Private test: IQ = 63.5 + 3 × 0.8823 × ((valid score − 3) × 35 / 14); maximum about 136.

A model's score is the average of its latest runs (up to 7), because a single run can swing by several points. Human IQ uses a mean of 100 and a standard deviation of 15, so the "higher than X% of people" figures above use that scale. Want to see the puzzle type for yourself? Try today's daily IQ puzzle.

Frequently asked questions

Which AI has the highest IQ?

On the private test that AI has never seen, GPT 6 Astra Ultra (OpenAI) leads with 133. On the public Mensa Norway test, GPT 6 Astra Ultra averages 148, close to the test's ceiling of 151. Scores as of 1 October 2026.

What is ChatGPT's IQ?

ChatGPT runs on OpenAI's GPT models. The strongest current one, GPT 6 Astra Ultra scores 133 on the private test and 148 on the public one. That private score is higher than about 98.6% of people.

What is Claude's IQ?

Anthropic's strongest current model, Claude-5.5 Opus MAX scores 133 on the private test and 144 on the public one.

What are Gemini's and Grok's IQ?

Google's Gemini 3.8 Flash scores 129 on the private test and 142 on the public one. xAI's Grok 4.7 High scores 129 on the private test and 142 on the public one.

Are you smarter than ChatGPT?

On these puzzles, only if your IQ is well above 125. The best models now beat about 97% of people on a private matrix test. But an IQ score measures one narrow skill. Models still make mistakes no person would, and they have no time limit, nerves or fatigue.

Why is the private test score lower than the public one?

The public Mensa Norway puzzles are online, so models may have seen them during training. The private test has never been published. Most models score 10 to 20 points lower on it, which suggests part of the public score is memory, not reasoning. The private scale also tops out at about 136, so the very best models hit its ceiling.

How is an AI IQ measured?

TrackingAI gives each model the same visual matrix puzzles humans take, every week. Text-only models get the puzzles described in words; vision models see the images. Correct answers are turned into an IQ: each one is worth about 3 points, and guessing at random gives about 63.5. A model's score is the average of its last 7 runs.

Is an AI IQ the same as a human IQ?

Not quite. The puzzles are the same, but the conditions are not: models have no time pressure, can be trained on similar puzzles, and the conversion is a simple mapping rather than a test normed on thousands of people. Read the scores as a useful comparison, not a like-for-like IQ.

How fast is AI getting smarter?

Very fast. The best single public-test run went from about 100 in 2024 Q2 to the 151 maximum by 2026. On the private test the best run went from about 97 to its 136 ceiling.

Could you beat the best AI?

Take 25 visual matrix puzzles, the same kind the models are tested on. About 12 minutes, no sign-up to start. Our test and the AI tests are different tests, so treat any comparison as a rough guide.

Take the IQ test

25 questions · about 12 minutes

Scores are TrackingAI's, averaged over each model's latest runs (up to 7) and rounded. Where a model has text and vision versions, the better private-test result is shown. We refresh this page as new results come in.