SGI-Bench — leaderboard
Scientific General Intelligence benchmark testing LLMs across deep research, idea generation, dry/wet experiments, and experimental reasoning. Evaluates scientific discovery capabilities.
Metric: SGI-Score. Source: github.com. Status: saturation imminent. 19 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 33.83 |
| 2 | Claude Sonnet 4.5 | 32.16 |
| 3 | Qwen 3 Max | 31.97 |
| 4 | GPT-4.1 | 31.45 |
| 5 | GPT-5.2 Pro | 31.09 |
| 6 | GPT-5 | 30.84 |
| 7 | O3 | 30.68 |
| 8 | Claude Opus 4.1 | 30.42 |
| 9 | O4 Mini | 30.14 |
| 10 | GPT-5.1 | 29.31 |
| 11 | Grok 4 | 28.68 |
| 12 | Qwen 3 VL 235B A22B | 28.32 |
| 13 | Gemini 2.5 Pro | 28.17 |
| 14 | Intern-S1 | 28.1 |
| 15 | GPT-4o | 26.87 |
Interactive version: theaggregate.ai/benchmark?slug=sgi-bench · How the rankings work · Data refreshed daily, snapshot 2026-07-22.