AI vs the world’s best teen mathematicians

01 Key findings

1. Gemini Deep Think: joint 27th of 630. Its IMO-certified 35 of 42 at the 2025 Olympiad equalled the lowest gold score. 26 contestants scored more, 46 scored the same and 558 (88.6%) scored less.IQ Metrics analysis of the IMO 2025 individual results and Google DeepMind, 2025
2. The problem no AI could do. In 2025 Gemini, OpenAI’s model, ByteDance’s Seed Prover and Harmonic’s Aristotle all failed Problem 6. Only 6 of 630 human contestants (1%) solved it in full.IMO 2025 individual results; the four labs' reports
3. AI’s blind spots are not human ones. In 2024 AlphaProof solved Problem 6, which only 5 of 609 humans solved, but missed Problem 5, which 154 humans solved. Its 28 points, with up to three days per problem, ranked joint 59th of 609: one point short of gold.IMO 2024 individual results; Google DeepMind, 2024
4. 2026: claims of 42 out of 42. Xiaohongshu and Huawei said their models scored full marks, and an informal test gave four more models 42. Only 7 of 666 human contestants did that. The IMO has published no statement on the AI entries.China Daily and AFP, July 2026; IMO 2026 edition page
5. Olympiad medallists become top mathematicians. 16 of the 38 Fields Medallists since 1990 (42.1%) won an IMO medal at school, 13 of them gold, including two of the four 2026 medallists.IMU Fields Medal list; IMO contestant records
6. Girls are 1 in 9 contestants, as in 1959. 76 of the 666 contestants in 2026 were girls (11.4%), the same share as at the first Olympiad (6 of 52, 11.5%). The low point was the 1970s, at 2%.IMO editions table, 1959-2026

02 The short answer in four numbers

AI now matches the best teenagers on points, but not on the same problems. The question it missed in 2025 was one that only six humans solved.

Leon Falkner, IQ Metrics Research

03 Where would AI have ranked among the world's best teenage mathematicians?

The International Mathematical Olympiad (IMO) sets national teams of up to six students ^[4] six problems in two sessions of 4.5 hours ^[10]. Every problem is worth 7 points, so the maximum is 42 ^[8]. Since 2024, AI labs have tried the same problems. We placed each reported AI score among the human totals of that year ^[3].

Figure 1

Every contestant at the last three Olympiads, and where AI landed

One dot per human contestant, stacked at their total out of 42 (609 in 2024, 630 in 2025, 666 in 2026). Dashed lines: the lowest score for gold, silver and bronze; dots coloured by medal. Navy and purple marks: AI results (AlphaProof was marked by graders DeepMind chose and took up to three days; Gemini Deep Think was certified by the IMO; OpenAI's and Seed Prover's results were reported by the labs; the 2026 scores of 42 are company claims and an informal test, not confirmed by the IMO).

GoldSilverBronzeone dot = one contestant

2024609 contestantsAlphaProof 28

Gold 29Silver 22Bronze 16

2025630 contestantsGemini Deep Think, OpenAI 35Seed Prover 30

Gold 35Silver 28Bronze 19

2026666 contestantsAI claims (companies, informal test) 42

Gold 29Silver 23Bronze 16

071421283542

2024. Google DeepMind’s AlphaProof and AlphaGeometry 2 scored 28 ^[8]. People first translated the problems into a formal language, and the systems took up to three days. 28 would have ranked joint 59th of 609: all 58 gold medallists scored more, and 17 humans scored exactly 28.

2025. Gemini Deep Think scored 35 within the 4.5-hour limit, and IMO coordinators graded and certified it ^[9]. OpenAI reported the same 35 for an experimental model, marked by three former medallists it chose ^[10]. 35 was exactly the gold cut-off: 26 humans scored more and 46 scored 35. ByteDance’s Seed Prover reported 30 points, confirmed with the IMO by its own account, after three days of work ^[11]: joint 100th of 630.

2026. Xiaohongshu said IMO graders gave its model dots-note-3.0 7 points on every problem ^[13], and Huawei reported full marks for Celia ^[14]. An investor’s informal test gave four other models 42 ^[14]. 7 of the 666 human contestants also scored 42. The IMO’s site carries no statement on these entries (news page read on 29 October 2026), so we label them company claims.

Table 1

Every AI result at the Olympiad, 2024-2026, problem by problem

Under each name: who marked it, and the place it would have taken. Green: full marks on that problem; red: no points; blue: part marks. Place: the rank the IMO would print for that total that year (of 609, 630 and 666 contestants).

AI system	Points	P1	P2	P3	P4	P5	P6
2024 AlphaProof (DeepMind) Graders the lab chose; would rank joint 59th	28	7	7	0	7	0	7
2025 Gemini Deep Think (DeepMind) IMO (confirmed); would rank joint 27th	35	7	7	7	7	7	0
2025 OpenAI model (OpenAI) Graders the lab chose; would rank joint 27th	35	7	7	7	7	7	0
2025 Seed Prover (ByteDance) IMO, says the company; would rank joint 100th	30	2	7	7	7	7	0
2025 Aristotle (Harmonic) Proof checker, no points	–	Proof	Proof	Proof	Proof	Proof	0
2026 dots-note-3.0 (Xiaohongshu) IMO, says the company; would rank joint 1st	42	7	7	7	7	7	7
2026 Huawei Celia (Huawei) IMO, says the company; would rank joint 1st	42	7	7	7	7	7	7
2026 4 frontier models (Deedy Das) Informal test; would rank joint 1st	42	7	7	7	7	7	7

04 Which Olympiad problems could AI not solve?

Points hide where the difficulty sits. In 2025 every reported AI system solved Problems 1 to 5 (Seed Prover only partly solved Problem 1) and none solved Problem 6, which 6 of 630 humans solved in full. In 2024 the pattern ran the other way: AlphaProof solved Problem 6, the hardest for humans (5 of 609), but not Problem 5, which a quarter of the contestants solved (154).

Figure 2

How many humans solved each problem in full, 2024-2026

Share of contestants with all 7 points on each problem. Every AI missed Problem 6 in 2025 (6 of 630 humans solved it); in 2024 AlphaProof solved Problem 6 (5 humans did) but not Problem 5 (154 humans did). Darker = more people solved it.

0%70%

P11P22P33P44P55P66202467.8%25.6%1.3%64.5%25.3%0.8%202558.4%40.2%16.2%54.3%34.1%1.0%202666.5%15.0%2.0%50.9%22.7%3.5%

The two “hard” slots, Problems 3 and 6, are where human scores collapse: in the three years, between 0.8% and 16.2% of contestants solved them in full, against 50.9% to 67.8% for Problems 1 and 4.

05 How many points does it take to win an Olympiad gold medal?

About one contestant in twelve wins gold: 1,917 of all 23,003 contestant places since 1959 (8.3%). The cut-off moves with the problems. Since 1982 gold has needed a median 31 of 42 points, from 24 in 2021 to 42 in 1987, when 22 contestants had perfect scores ^[2].

Figure 3

Points needed for a medal, 1982-2026

The lowest total that won gold, silver and bronze each year since 1982 (six problems worth 7 points, 42 in all, since 1981). Gold needed between 24 (2021) and 42 (1987); the median is 31.

GoldSilverBronze

71421283542

292316

19821993200420152026

Figure 4

Perfect scores at the Olympiad, 1959-2026

Contestants with full marks each year (314 in all). No Olympiad was held in 1980. Before 1981 the maximum score changed from year to year.

1

01960

1

1

0

1

2

11

5

16

3

31970

1

8

1

6

6

1

5

1

4

–1980

26

3

4

8

2

3

22

5

10

41990

9

4

2

22

14

1

4

1

0

42000

4

3

3

4

16

3

0

3

2

12010

1

1

0

3

1

6

0

2

6

12020

1

10

5

1

5

72026

In 67 Olympiads, 23,003 contestant places have produced 314 perfect scores and 1,917 gold medals ^[1]. The contest grew from 52 students from 7 countries in 1959 to 666 from 117 in 2026.

06 How many girls compete at the International Math Olympiad?

Few, and no more than at the start. At the first Olympiad in Romania in 1959, 6 of the 52 contestants were girls (11.5%). In 2026 it was 76 of 666 (11.4%) ^[1]. The share fell to 2% in the 1970s and has climbed back in every decade since.

Figure 5

Girls among Olympiad contestants, by decade

Share of contestant places held by girls, as recorded by the IMO (one person who took part three times counts three times). Girls' share: 1959 11.5%, 1960s 5.1%, 1970s 2%, 1980s 3.9%, 1990s 6.4%, 2000s 8.2%, 2010s 10%, 2020-26 11.1%.

11.5%1959

5.1%1960s

2.0%1970s

3.9%1980s

6.4%1990s

8.2%2000s

10.0%2010s

11.1%2020-26

07 Which countries win the International Math Olympiad?

China, mostly. It has topped the unofficial team ranking 19 times since 2000 and in 2026 again sent six gold medallists ^[4]. The United States came first in 1977, 1981, 1986, 1994, 2015, 2016, 2018, 2019, 2024. The ranking is unofficial: the IMO is a contest for individuals.

Figure 6

Team rank of seven leading countries, every fourth year, 2010-2026

Unofficial team ranking by the sum of the six team members' points (1 = top). Russia has no rank in 2022-2025, when it did not compete as a team. Hover a line for every round.

United StatesChinaOther countries

2010'102014'142018'182022'222026'26

China 1United States 3South Korea 4Russia 2Singapore 22Japan 7United Kingdom 25

1 China2 United States6 South Korea3 Russia4 Singapore10 Japan8 United Kingdom

08 Do Math Olympiad medallists become great mathematicians?

Often. The Fields Medal is the top award in mathematics, given every four years to up to four mathematicians under 40 ^{[6] [7]}. Of the 38 medallists since 1990, 16 had won an Olympiad medal at school, 13 of them gold ^[5]. The IMO noted that two of the four 2026 medallists, Yu Deng and Jacob Tsimerman, are former contestants ^[6].

Table 2

Fields Medallists who won an Olympiad medal at school

Fields Medal	Mathematician	IMO team	Olympiad results	Best medal
2026	Jacob Tsimerman	Canada	2003 Gold (30/42), 2004 Gold (42/42)	Gold
2026	Yu Deng	China	2006 Gold (35/42)	Gold
2018	Akshay Venkatesh	Australia	1994 Bronze (28/42)	Bronze
2018	Peter Scholze	Germany	2004 Silver (31/42), 2005 Gold (42/42), 2006 Gold (35/42), 2007 Gold (36/42)	Gold
2014	Artur Avila	Brazil	1995 Gold (37/42)	Gold
2014	Maryam Mirzakhani	Iran	1994 Gold (41/42), 1995 Gold (42/42)	Gold
2010	Elon Lindenstrauss	Israel	1988 Bronze (17/42)	Bronze
2010	Ngô Bảo Châu	Vietnam	1988 Gold (42/42), 1989 Gold (40/42)	Gold
2010	Stanislav Smirnov	USSR	1986 Gold (42/42), 1987 Gold (42/42)	Gold
2006	Grigori Perelman	USSR	1982 Gold (42/42)	Gold
2006	Terence Tao	Australia	1986 Bronze (19/42), 1987 Silver (40/42), 1988 Gold (34/42)	Gold
2002	Laurent Lafforgue	France	1984 Silver (27/42), 1985 Silver (25/42)	Silver
1998	Richard Borcherds	United Kingdom	1977 Silver (29 pts), 1978 Gold (39 pts)	Gold
1998	Timothy Gowers	United Kingdom	1981 Gold (42/42)	Gold
1994	Jean-Christophe Yoccoz	France	1973 Silver (32 pts), 1974 Gold (40 pts)	Gold
1994	Pierre-Louis Lions	France	1973 no medal (9 pts)	Bronze
1990	Vladimir Drinfeld	USSR	1969 Gold (40 pts)	Gold
1978	Grigory Margulis	USSR	1962 Silver (36 pts)	Silver

The other 22 medallists since 1990 have no Olympiad medal on record: the Olympiad is one road to the top, not the only one.

09 What does AI's Olympiad score say about intelligence?

Olympiad problems are new every year, so they cannot be memorised, which is why AI labs use them to compare their systems with people. But the comparison is narrow. The students had 9 hours and no tools; some AI systems had days or human-made formal translations, and the labs chose how to report them. Matching the gold cut-off on six problems is not the same as general intelligence, just as a high score on one IQ subtest is not a full IQ. For a short reasoning test of the kind people take, see our IQ test; for how one famous puzzle trips up even strong reasoners, see the world’s shortest IQ test.

10 How we built the data

1. 01

   Take the yearly facts

   For every Olympiad from 1959 to 2026 we read the IMO’s edition pages: countries, contestants, girls, medals, medal cut-offs and perfect scores ^{[1] [2]}.

2. 02

   Count the 2024-2026 scores

   From the individual results of 2024, 2025 and 2026 we counted how many contestants had each total and full marks on each problem ^[3]. We keep counts only, no names; the counts match the IMO’s medal and perfect-score totals exactly.

3. 03

   Collect every AI result

   Each lab’s own report, read at the source: score, problems solved, time and who marked it ^{[8] [9] [10] [11] [12] [13] [14]}.

4. 04

   Place AI among the humans

   For each AI score: humans above, equal and below, and the rank the IMO would print.

5. 05

   Match the Fields Medallists

   All 68 Fields Medallists ^[7] searched in the IMO’s contestant records by name, country and age ^[5]. 18 names matched; one (Pierre-Louis Lions, France 1973, no medal) is not confirmed as the same person and is left out.

6. 06

   Rank the teams

   Team ranks per year for eight leading countries, from the IMO’s country pages ^[4].