Can AI solve math olympiad or research-level math problems?
YesYes. AI systems reached gold-medal level at the 2025 International Mathematical Olympiad, several claimed perfect scores in 2026, and AI has resolved some genuinely open research problems, though most so far are easier ones and claims still need expert checking.
First shown: Jul 21, 2025. Google DeepMind's Gemini Deep Think solved 5 of 6 IMO 2025 problems in natural language within the time limit, scoring 35/42, gold-medal level, as graded by IMO coordinators; OpenAI reported the same score two days earlier under its own grading. (Google DeepMind)
What it means
The International Mathematical Olympiad is the hardest math contest for school students: six problems over two 4.5-hour sessions, answered with written proofs. In July 2025 an experimental OpenAI model and Google DeepMind's Gemini Deep Think each solved five of six problems for 35 of 42 points, gold-medal level, writing proofs in ordinary language. IMO coordinators graded Google's answers, though they did not vouch for the system itself; OpenAI's were graded by three former medalists.
A year earlier Google's AlphaProof had reached silver, but only after humans translated the problems into a formal language and with two to three days of computing. At the 2026 olympiad in Shanghai, Huawei and Xiaohongshu said their models scored a perfect 42/42, with Xiaohongshu saying its answers went through official IMO judging; seven of 666 human contestants did the same.
Research math is harder, because nobody knows the answer in advance. In October 2025 an OpenAI executive claimed GPT-5 had solved ten open Erdos problems, but it had found solutions already in the literature, and the post was deleted. In January 2026 Terence Tao confirmed GPT-5.2 Pro had solved Erdos problem #728 more or less on its own, checked by computer, while estimating that only about 1-2% of open Erdos problems are easy enough for today's AI.
Bigger results followed. In May 2026 OpenAI said an internal model disproved a 1946 Erdos conjecture about distances between points, with outside mathematicians checking the proof. In July 2026 Anthropic researcher Levent Alpoge announced a counterexample to the 1939 Jacobian conjecture that he credits to Claude; anyone can check the formula, but it had not been peer reviewed.
The evidence ladder
Every capability on this site is graded on the same five steps, from a single demonstration to real-world consequences.
| Step | Status | When | Evidence |
|---|---|---|---|
| Demonstrated shown at least once | yes | Jul 21, 2025 | Gemini Deep Think scored 35/42 at IMO 2025 in natural language within the time limits; IMO coordinators confirmed the proofs but did not validate the system. (Google DeepMind) |
| Reproduced independently repeated | yes | Jul 23, 2026 | At IMO 2026 Huawei and Xiaohongshu said their models scored 42/42; a venture investor's own test found four other models also scored 42/42, without official grading. (AFP via The Standard) |
| Accessible available to ordinary people, or criminals | partly | Aug 1, 2025 | A faster Gemini Deep Think, bronze-level on IMO 2025 by Google's own tests, went to paying AI Ultra subscribers; the gold-level version went only to selected mathematicians. (Google) |
| Reliable works most of the time for a defined task | partly | Jan 16, 2026 | Tao confirmed an AI solution to Erdos problem #728 but estimated only about 1-2% of open Erdos problems are within reach of current AI with little help; an October 2025 claim of ten solutions collapsed. (The Decoder) |
| Consequences real-world use or harm documented | yes | Jul 20, 2026 | A counterexample to the 1939 Jacobian conjecture, credited to Claude, followed OpenAI's May 2026 disproof of a 1946 Erdos conjecture; the Jacobian result was not yet peer reviewed. (Forklog) |
From the record
- OpenAI said an experimental model scored 35/42 at IMO 2025 under contest conditions, graded by three former medalists. (Simon Willison)
- An OpenAI executive deleted a claim that GPT-5 had solved ten open Erdos problems after it emerged that the model had found existing published solutions. (TechCrunch)
- OpenAI said an internal model disproved Erdos's 1946 unit distance conjecture, with external mathematicians checking the proof. (OpenAI)
- A written-up counterexample to the 1939 Jacobian conjecture was posted, credited to Anthropic's Claude after Levent Alpoge announced it. (ulam.ai)
What it means for you
Solving hard, well-defined math problems is no longer something only top humans can do. Students and teachers face AI that can do contest math, and mathematicians are starting to use it as a collaborator, though its claims still need checking.
What to do
- Check AI math before relying on itAsk for the steps and check them, or use a tool that verifies proofs; even good models make confident mistakes.
- Teachers: test understandingContest-level homework can now be solved by AI; oral explanations and in-class work show what students know.
- Treat breakthrough claims with careWait for mathematicians to check a claimed AI result; the October 2025 Erdos claim fell apart within days.
- Use it to learn, not to skip learningAsking AI to explain a solution can help; copying it teaches nothing.
Questions people ask
Did AI really win a gold medal at the Math Olympiad?
AI did not compete officially, but in 2025 two systems scored 35/42, gold-medal level, on the same problems under the same time limits.
Has AI solved an unsolved math problem?
Yes. Terence Tao confirmed an AI solution to Erdos problem #728 in January 2026, and AI was credited with disproving older conjectures in May and July 2026.
Will AI replace mathematicians?
Not yet. Tao estimates only about 1-2% of open Erdos problems are within reach of current AI, and harder work still needs human strategy.
Can I trust a chatbot's math answer?
Not blindly. Even strong models make errors, so check the steps; consumer versions are often weaker than the systems used in these results.
Related
Sources
- Google DeepMind: Gemini Deep Think IMO gold (Jul 2025)
- Simon Willison on OpenAI's IMO result (Jul 2025)
- Google: Gemini 2.5 Deep Think release (Aug 2025)
- TechCrunch: OpenAI's embarrassing math (Oct 2025)
- The Decoder: Tao on GPT-5.2 Pro and Erdos #728 (Jan 2026)
- OpenAI: model disproves discrete geometry conjecture (May 2026)
- AFP via The Standard: AI scores 100% at IMO 2026
- Forklog: Claude and the Jacobian conjecture (Jul 2026)
- A Counterexample to the Jacobian Conjecture (ulam.ai)
Spot an error or something out of date? Tell us; corrections are made and logged. Written with AI assistance from public sources; every fact checked and approved by Colin Balfe. General information, not legal, medical or financial advice.