SPM-Bench: leaderboard
Metric: Strict Imperfection Penalty F1 (0-1, shown times 100): 1 for an exact option set, otherwise 0.6 times an F1 whose precision weights each wrongly selected option six times, over 2,703 PhD-level multi-select image questions on scanning probe and electron microscopy (STM, AFM, SEM, TEM) synthesized from 2023-2025 papers; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 93.6 | #45 |
| 2 | Claude Sonnet 4.6 | 90.8 | #85 |
| 3 | Qwen 3.5 Plus | 88.1 | #123 |
| 4 | Claude Opus 4.6 | 86.3 | #60 |
| 5 | Kimi K2.5 | 85.5 | #139 |
| 6 | GPT-5.2 | 81.4 | #105 |
| 7 | GPT-5.1 | 80.5 | #131 |
| 8 | Seed 2.0 Pro | 79.1 | #111 |
| 9 | Qwen 3.6 Plus | 79 | #113 |
| 10 | Gemini 3.1 Pro (Preview) | 76.2 | #54 |
| 11 | Claude Opus 4.5 | 74.5 | #79 |
| 12 | Gemini 3 Flash (Preview) | 71 | #78 |
| 13 | Qwen 3 VL 235B A22B (Thinking) | 67.6 | #228 (Qwen 3 VL 235B A22B) |
| 14 | GPT-5.5 | 59.4 | #26 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=spm-bench · How It Works · Data refreshed daily, snapshot 2026-10-11.