Rabdos Math Bench
Twenty-five original problems, kept off the public web.
Results
Last updated 07/12/2026
| Rank | Provider | Model | Score |
|---|---|---|---|
| 1 | OpenAI | GPT-5.5 (xhigh) | 75% |
| 1 | OpenAI | GPT-5.6 Sol (max) | 75% |
| 3 | Anthropic | Claude Opus 4.8 (max) | 72.5% |
| 4 | Anthropic | Claude Fable 5 (max) | 67.5% |
| 5 | Gemini 3.1 Pro (high) | 35% | |
| 6 | Gemini 3.5 Flash (high) | 25% |
Each problem is attempted multiple times; scores are the fraction of correct attempts.
Rabdos Math Bench is a held-out benchmark for advanced mathematical reasoning in frontier AI models. Its problems are written for the benchmark and unpublished, to prevent contamination. Each resolves to a single numeric answer, so a model's output is checked exactly.
Conditions on a Problem
Every problem in the benchmark satisfies five conditions.
- Original. Written for this benchmark. Never published, and kept off the public web.
- Diverse. Drawn from across mathematics, not from a single field.
- Determinate. One fixed numeric answer, and a statement screened for ambiguity.
- Verified. Checked in full by mathematicians before admission.
- Earned. Answers come from spaces too large to guess; every problem is screened for shortcuts.
Three of the twenty-five
Contact
For model evaluations, research access, or methodology details, contact the Rabdos team.