
On this page
Is AI smarter than humans? In 2026 the honest answer is “at some things, by a wide margin, and at others, not even close.” This page is a scorecard: the main tests where AI and humans have been measured on the same tasks, with sources, and the places where people still win.
Short answer: AI is smarter than most humans on exams, maths competitions and reasoning puzzles. It reached gold-medal level at the 2025 Maths Olympiad and scored 95% on the ARC-AGI-2 puzzle test in 2026. But humans still win at finishing real computer tasks, physical work and some simple perception. Overall, AI is not yet smarter than humans.
The scorecard: AI vs humans in 2026
Short answer: AI leads on tests with clear right answers; humans lead on open-ended, physical and long tasks.
| Test | Best AI | Humans | Who wins |
|---|---|---|---|
| Maths Olympiad (IMO 2025) | Gemini Deep Think: 35/42, gold (officially graded) | Gold is reached only by top contestants | AI ≈ elite humans |
| ARC-AGI-2 (novel puzzles) | GPT-6: 95% (Sept 2026) | Every task solved by at least 2 people in ≤2 tries | AI |
| Humanity’s Last Exam (expert questions) | +30 points in one year (2026 AI Index) | Written by experts to favour experts | Narrowing fast |
| Mensa Norway IQ test | OpenAI o3: 136 (Apr 2025) | Average 100 | AI (with caveats, see AI IQ) |
| OSWorld (using a computer) | 66.3% | 72.35% | Humans |
| ClockBench (reading analog clocks) | 50.6% | ~90% | Humans |
| Real household tasks (robots) | 12% success | Routine for adults | Humans |
| Structured agent tasks | Fail about 1 in 3 | — | Humans (reliability) |
Sources: Stanford HAI 2026 AI Index, ARC Prize, Euronews, CryptoSlate/TrackingAI. Links below.
Is AI smarter than humans at maths and reasoning?
Short answer: yes, at competition-style problems.
At the 2025 International Mathematical Olympiad, Google DeepMind’s Gemini Deep Think solved five of six problems for 35 points, a gold score, working in natural language within the 4.5-hour limit; its result was officially certified. OpenAI reported the same score with an experimental model, checked by former medalists rather than the IMO. A year earlier, DeepMind’s system had scored silver.
On ARC-AGI-2, a test of learning a new rule from a few examples, the ARC Prize leaderboard shows GPT-6 at 95% in September 2026. The benchmark was built so that every task had been solved by at least two people in two attempts or fewer, in testing with 400 participants.
Is AI smarter than humans at exams and knowledge?
Short answer: it knows far more, and now scores near expert level on hard exams.
No person can match a language model’s breadth of recall. Even Humanity’s Last Exam, built from questions expert humans could answer but AI could not, saw a 30-point jump in a single year, according to Stanford. The AI Index’s wider finding is that tests designed to last years are now “saturated in months.”
Where do humans still beat AI?
Short answer: doing whole jobs, the physical world and some basic perception.
- Finishing tasks. On OSWorld, which asks AI to use real software, the best system reached 66.3% vs a 72.35% human baseline.
- Reliability. AI agents still fail about one in three structured tasks. A person who failed a third of routine tasks wouldn’t keep their job.
- Bodies. Robots succeed in just 12% of real household tasks.
- Strange gaps. The top model in 2026 read analog clocks correctly only 50.6% of the time.
Researchers call this jagged intelligence. It’s why a system can win maths gold and still fumble a clock face. We look at the deeper reasons, such as learning efficiency, embodiment and energy, in AI vs human intelligence.
Does “smarter” even mean the same thing?
Short answer: not quite. Benchmarks measure narrow skills, not intelligence as a whole.
A test score shows that AI can do that test. Humans who score well on one test usually do well on many related tasks; AI’s results transfer less predictably. That’s also why IQ scores for chatbots are a weak signal (see AI IQ). Some benchmark questions are flawed too: the AI Index found invalid-question rates as high as 42% on one popular test.
When will AI be smarter than humans overall?
Short answer: forecasts range from 2026 to the 2040s.
Elon Musk said at Davos in January 2026 that AI could be smarter than any single human by the end of 2026 or 2027. A 2023 survey of 2,778 AI researchers gave a 50% chance that machines would beat humans at every task by 2047. Our how close are we to AGI page compares every major prediction. Beyond that point lies superintelligence: AI far smarter than all of us.
Our take
AI is already smarter than humans at answering hard questions, and not yet smarter at doing things. The shift to watch isn’t the next benchmark record; it’s agents and robots closing the reliability gap. When the bottom rows of the scorecard flip, “is AI smarter than humans?” will have a simple answer.
Enjoyed this? Get the SI Briefing
One short email a week: what changed on the road to super intelligence, and why it matters. No spam, unsubscribe anytime.
Frequently asked questions
Is AI smarter than humans?
On many tests, yes: AI beats most people at maths competitions, puzzle benchmarks and exam questions. Overall, no: humans still do better at completing real computer tasks, physical work and some simple perception tasks, and AI makes errors a person wouldn't.
Is AI smarter than the smartest human?
Only in narrow areas. AI matched gold-medal level at the 2025 International Mathematical Olympiad, where only a small share of the world's best young mathematicians win gold, but no AI yet outperforms top experts across their whole job.
When will AI be smarter than humans?
Forecasts range from 2026–2027 (Elon Musk at Davos 2026) to 2047 (median of a 2023 survey of AI researchers for beating humans at every task). See our AGI timeline page for every major prediction.
Can AI outsmart humans?
In specific, well-defined tasks it already does. Outsmarting people in open-ended real-world situations is harder, because today's AI still struggles with long tasks, reliability and the physical world.
What can humans do that AI can't?
In 2026, humans still clearly beat AI at physical household tasks, finishing long multi-step computer jobs reliably, and some everyday perception like reading analog clocks. Humans also learn new skills from far fewer examples.
6 Sources
- Stanford HAI, 2026 AI Index: Technical performance
- ARC Prize, ARC-AGI results
- ARC Prize, ARC-AGI-2 technical report (human testing)
- Euronews: Did Google DeepMind or OpenAI win gold at the IMO? (22 Jul 2025)
- CryptoSlate: OpenAI's o3 scores 136 on Mensa Norway test (Apr 2025, TrackingAI data)
- AI Magazine: Musk on AI and superintelligence at Davos 2026