On an IQ-style quiz never published online, top AI scores rose from 85 to 136
The quiz is kept offline so models can't have trained on it. Scores rose about 50 points since May 2024, but an AI ‘IQ’ means little.
Top AI score on an offline IQ-style test
Best monthly score on TrackingAI's verbal IQ-style quiz that has never been published online. Human average = 100.
What this measures: A short IQ-style pattern quiz that has never been published online, so models can't have memorised it. 100 is the human average.
- Top model (offline test)
- Reference level
Monthly maxima, dated to the end of each month. IQ norms are designed for humans and aren't validated for AI; this is a small matrix-style quiz, not a measure of general intelligence. On the public Mensa Norway test (which may be in training data) scores are higher, about 150.
Show data table and sources
| Date | Label | Value |
|---|---|---|
| 31 May 2024 | Claude 3 | 85.5 points |
| 30 Nov 2024 | o1 | 96.5 points |
| 30 Nov 2025 | Grok 4 | 123 points |
| 31 May 2026 | GPT-5 Pro | 127 points |
| 31 Oct 2026 | GPT Astra Ultra (TrackingAI label) | 136 points |
| Human average (100) | 100 points |
- Top AI IQ scores over time (TrackingAI)
- Offline-test monthly top scores (CSV) (TrackingAI)
People reach for IQ as shorthand for how clever a model is, so here is the best available data of that kind, with its limits. TrackingAI gives models two tests. One is the public Mensa Norway quiz. The other was written by a Mensa member and never published online, so a model cannot have memorised it from the web.
On the unpublished quiz, the best score of the month rose from 85.5 (Claude 3, May 2024) to 96.5 (o1, November 2024), 123 (Grok 4, November 2025) and 136 (October 2026). On the public quiz, scores run higher, around 150, which is what you would expect if the questions have been seen in training.
The gap between the two is the useful part. It shows how much of a high score can come from having met the questions before rather than from working them out.
What it means
Scores climbed by about 50 points in two and a half years on a quiz models cannot have memorised, so the improvement is real and not just familiarity with the questions. It does not mean a model is as capable as a person who scores 136: these are short pattern puzzles, and human IQ scales were never validated for machines. When you see an IQ figure quoted for an AI, ask whether the test was public.
Caveats
- Human IQ norms were never validated for AI, so a score here is not a measure of general intelligence.
- These are short pattern puzzles, and TrackingAI publishes little about its method (how many questions, how scores are converted).
- TrackingAI’s model names do not always map to a specific released version, and the best score of a month moves around.