IQ to Percentile Conversion
Here's how IQ scores translate to percentiles, giving you a sense of where an AI system stands relative to human performance:
| IQ Score | Percentile | Description |
|---|---|---|
| 70 | 2nd | Significantly below average |
| 85 | 16th | Below average |
| 100 | 50th | Average |
| 115 | 84th | Above average |
| 130 | 98th | Highly gifted |
| 145 | 99.9th | Exceptionally gifted |
| 160 | 99.99th | Profoundly gifted |
Remember: When an AI achieves "130 IQ," it means it outperformed 98% of humans on that specific test. The same AI might perform at 70 IQ on a different type of reasoning task.
Frontier AI IQ — June 2026
aiiq.org maps 12 hard benchmarks (FrontierMath, ARC-AGI-2, GPQA Diamond, SWE-bench, Humanity's Last Exam...) onto the human IQ scale. The frontier as of June 10, 2026:
| Model | Est. IQ | Human percentile |
|---|---|---|
| GPT-5.5 | ~130–136 | 98th–99th |
| Claude Opus 4.8 | ~130 | 98th |
| Gemini 3.1 Pro | ~127 | 96th |
| Claude Opus 4.7 | ~126 | 96th |
| Kimi K2.6 | ~122 | 93rd |
| GLM-5.1 | ~121 | 92nd |
| Qwen 3.7 Max | ~118 | 88th |
| Grok 4.3 | ~117 | 87th |
- • One year ago the frontier was o3 at ~112. That's ≈ +1.3 IQ points per month, measured.
- • On novel, offline tests scores drop 20–40 points — public-test results partly reflect memorization.
- • The jagged truth: the same models solving Erdős problems still fumble tasks a child finds trivial. IQ is one lens, not the territory.
- • Live trackers: aiiq.org, trackingai.org, metr.org
The "Highest Human IQ" Question
- • Real IQ tests stop measuring around 160 (99.99th percentile) — beyond that there aren't enough humans to norm the test against.
- • Terence Tao's famous "220–230 IQ" is a ratio-IQ extrapolation from scoring 760 on the math SAT at age 8 — not a measured deviation IQ. Tao himself calls the figure noise and has claimed only "greater than 175."
- • Marilyn vos Savant's record 228 was a childhood ratio score; her adult deviation-test result was ≈186. Guinness retired the "Highest IQ" category in 1990 as statistically unreliable.
- • So when honest AI scores reach ~160, "AI IQ" stops being a number at all. That's why the IntExp chart marks everything above 160 as off the human scale — from there on, you need different rulers: task horizons, percentile-of-experts, things no human can do at any speed.
AI IQ Benchmarks & Progress
Click on any image below to view it fullscreen and explore the latest AI capability assessments:
GPQA Diamond Benchmark Progress
AI performance on graduate-level scientific questions (July 2023 to May 2025)
PhD-Level Scientific Reasoning
Benchmarking AI on doctoral-level scientific problems
AI IQ on Codeforces
Competitive programming performance showing AI reaching exceptional problem-solving levels
AI IQ Progress Tracking (11 Months)
MaximumTruth's tracking of AI IQ improvements over time - offline testing, not MENSA certified