SGI-Bench Deep Research: leaderboard
SGI-Bench subscore for deep scientific research tasks that require literature-grounded synthesis, multi-step reasoning, and domain expertise.
Metric: Deep Research Score. Source: github.com. Saturation forecast: Around 2030. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 18.48 | #80 |
| 2 | Intern-S1 | 15.74 | #285 |
| 3 | GPT-5.2 Pro | 15.72 | #61 |
| 4 | Qwen 3 Max | 15.38 | #208 |
| 5 | Gemini 2.5 Pro | 15.09 | #152 |
| 6 | GPT-5 | 14.47 | #98 |
| 7 | Claude Sonnet 4.5 | 13.84 | #142 |
| 8 | Grok 4 | 13.31 | #181 |
| 9 | Claude Opus 4.1 | 12.93 | #140 |
| 10 | O3 | 12.89 | #124 |
| 11 | Qwen 3 VL 235B A22B | 11.97 | #238 |
| 12 | O4 Mini | 11.95 | #180 |
| 13 | GPT-5.1 | 11.64 | #135 |
| 14 | GPT-4.1 | 11.32 | #248 |
| 15 | Gemini 2.5 Flash | 10.69 | #245 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=sgi-bench-deep-research · How It Works · Data refreshed daily, snapshot 2026-10-20.