RiemannBench
Math · 2026-04-08
RiemannBench evaluates unconstrained AI research agents on 25 research-level math problems with full access to coding tools, search, and open-ended reasoning. Scores average repeated attempts per problem; the 25-item count records distinct problems, not a single-run binomial denominator.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Carbon | 80.0 |
| Gemini 4 Argon | 76.0 |
| GPT-5.6 Sol | 74.4 |
| GPT-6 Astra | 72.0 |
| Claude Opus 5.5 | 69.6 |
| Claude Opus 5 | 68.0 |
| GPT-5.6 Terra | 66.4 |
| Claude Fable 5.1 | 65.6 |
| GPT-6 Sol | 64.8 |
| Claude Fable 5 | 60.0 |
| GPT-6 Luna | 59.2 |
| GPT-5.5 | 55.2 |
| Gemini 3.8 Flash | 51.2 |
| Opus 4.8 | 47.2 |
| GPT-5.4 | 41.6 |
| Gemini 3.7 Flash | 39.2 |
| Grok 4.5 | 38.4 |
| GPT-5.2 | 37.6 |
| Kimi K3 | 37.6 |
| Gemini 3.5 Flash | 36.8 |
| Gemini 3.1 Pro Preview | 33.6 |
| Opus 4.7 | 32.8 |
| Gemini 3.6 Flash | 30.4 |
| Muse Spark 1.3 | 28.0 |
| DeepSeek V4.1 Flash | 24.0 |
| DeepSeek V4 Flash Vision Exp | 23.2 |
| Muse Spark 1.2 | 23.2 |
| Muse Spark 1.1 | 20.8 |
| Inkling | 15.2 |
| Qwen3.7-Max (source unspecified) | 15.2 |
| Muse Glimmer 30B | 13.6 |
| GLM-5.3 | 12.0 |
| Kimi K2.5 | 12.0 |
| Gemini 3.5 Flash-Lite | 11.2 |
| GLM-5.2 | 10.4 |
| GLM-5.3 Flash | 8.8 |
| DeepSeek-V3.2 | 8.0 |
| Kimi K2.6 | 7.2 |
| DeepSeek-V4-Pro | 5.6 |