CMT-Benchmark: leaderboard
CMT-Benchmark — condensed-matter theory problems written by expert physicists (Hartree-Fock, exact diagonalization, DMRG and more); frontier models still score low.
Metric: Pass@1 (%). Source: arxiv.org. Saturation forecast: Around July 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 30 | #98 |
| 2 | O3 | 26 | #124 |
| 3 | GPT-5 Mini | 24 | #186 |
| 4 | O4 Mini | 18 | #180 |
| 5 | Gemini 2.5 Pro | 14 | #152 |
| 6 | GPT-5 Nano | 14 | #425 |
| 7 | Llama 4 Maverick | 12 | #464 |
| 8 | Gemini 2.0 Flash | 10 | #342 |
| 9 | Claude Opus 4.1 | 8 | #140 |
| 10 | Claude Sonnet 4 | 6 | #199 |
| 11 | Claude 3.7 Sonnet | 6 | #249 |
| 12 | Gemini 2.5 Flash | 4 | #245 |
| 13 | GPT-4.1 | 4 | #248 |
| 14 | DeepSeek V3 | 4 | #321 |
| 15 | GPT-4o | 2 | #343 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cmt-benchmark · How It Works · Data refreshed daily, snapshot 2026-10-20.