Wikipedia Unsolved Math Problems TrackerBeta

A controlled baseline of 457 diverse problems to calibrate the unprecedented progress of AI in mathematics.

AI progress in mathematics has been absolutely extraordinary. In a little over a year, frontier models went from barely achieving gold medals at the International Math Olympiad to a claimed resolution of the Navier–Stokes Millennium problem, one of the holy grails of mathematics.

Given this rapid rate of progress, it's critical to determine if AI is merely conquering “AI shaped” problems, or truly conquering mathematics altogether, replacing the role of human insight in proof generation.

Measuring this is difficult, since frontier labs don't report failures on open problems, only successes. We need a large, diverse set of open problems which are, with high probability, important to their fields. For this, we argue Wikipedia’s public List of unsolved problems in mathematics, snapshotted in January 2024, provides a particularly robust sample.

The dataset is in beta, and was generated using LLMs and has not been reviewed by mathematicians. See About the Data for more limitations and planned adjustments.

MIT FutureTech — Innovations that Shape the World
NoteOne day after this was launched, OpenAI released nearly 722 manuscripts, each making progress on an open problem. We are working on integrating these results ASAP. To a first approximation, it seems 30–40 additional problems on this list were resolved.
Corpus

Headline Numbers

457open problems
listed by Wikipedia in January 2024
30problems “resolved”
9 accepted, 21 still contested
57%involved an AI system
43%done entirely by AI
of the 30 resolutions
158saw any progress
23% of them involved an AI system

AI's Contribution, Over Time

Cumulative Resolutions

Every problem which has been resolved (or claimed to be resolved), stacked by how it was produced
01020304050Jan 2024Jul 2024Jan 2025Jul 2025Jan 2026Jul 2026Number of Problems Resolved
Human only 13With AI 4AI only 13

Cumulative Progress

Every problem which has seen a material advance – a case settled, a bound improved, or a claimed resolution – stacked by how it was produced
050100150Jan 2024Jul 2024Jan 2025Jul 2025Jan 2026Jul 2026Number of Problems Materially Advanced
Resolved, human only 13Resolved, with AI 4Resolved, AI only 13Progress, human only 107Progress, with AI 6Progress, AI only 14

AI's Contribution, by Problem Status

Every status category, split by who did the work

Hover any band for its count and share
Resolved9
Resolution claimed21
Progress of some kind108128
Still open278278
No AI involvementAI-assistedAI-generatedNot yet solved

Resolutions by Model Release Date

Problems credited, by when the system shipped

One bar per released system, at its release date; height is the problems it is credited on, split the same way as the bars below
024681012141618Apr 2025Jul 2025Oct 2025Jan 2026Apr 2026Jul 2026
AI-generated 38AI-assisted 18
A bar sits at the month a system first became available to the people who used it: the private preview where there was one, otherwise the public release. A problem credited to systems that arrived in different months appears once under each of them, so the bars sum to more than the number of problems. 8 systems are not on the chart at all, having no release or preview date to place them.

AI Systems Credited

Individual systems credited

Problems each named system contributed to, grouped by lab and split by kind of contribution
OpenAI 29 problems
GPT-6 Astra9211
GPT-5.6 Sol6410
GPT-5.6 Sol Pro44
OpenAI internal model33
GPT-522
GPT-5.5 Pro2
GPT-5.6 Sol Ultra22
ChatGPT (unknown version)1
Codex (unknown version)1
gpt-oss-120b1
gpt-oss-20b1
Anthropic 11 problems
Claude Fable 5347
Claude Opus 522
Anthropic internal model1
Claude (unknown version)1
Claude Fable 5.11
Claude Opus 4.61
Claude Sonnet 3.71
Google DeepMind 4 problems
AlphaEvolve22
Gemini 2.0 Flash1
Gemini 2.0 Pro1
Gemini 3 Pro1
Other 6 problems
AxiomProver1
Caltech RL system1
EinsteinArena1
Grok 4.61
PackingStar1
SimpleTES1
the Danus system1
TTT-Discover1
A problem can credit more than one system, so these bars sum to more than the number of problems involved. Proof assistants used for verification (Lean 4, Nanoda) are not counted as AI systems.