BenchTable - STEM: leaderboard
Dubesor's small-scale personal benchmark, category score for the 19 science, technology, engineering and mathematics tasks: difficulty-weighted points (pass, refine, fail, refusal) divided by the best points achieved on each task across models. Percent scale, higher is better; scores can be negative. The benchmark is archived, so the table no longer changes.
Metric: Weighted Score (%). Source: dubesor.de. Saturation forecast: Around March 2027. 347 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 (2024-12-17) | 93.8 | #157 |
| 2 | Claude Sonnet 4 | 93.2 | #207 |
| 3 | DeepSeek R1 0528 | 92.9 | #232 |
| 4 | Claude Opus 4.1 | 90.9 | #146 |
| 5 | Claude Opus 4.1 (Thinking) | 89.6 | #146 (Claude Opus 4.1) |
| 6 | GPT-5 | 87.9 | #105 |
| 7 | Claude Opus 4 (Thinking) | 87.2 | #170 (Claude Opus 4) |
| 8 | DeepSeek V3.2 Speciale | 87.1 | #138 |
| 9 | Gemini 3.1 Pro (Preview) | 86.6 | #64 |
| 10 | Claude Opus 4.7 | 86.1 | #51 |
| 11 | Claude 3.7 Sonnet (Thinking) | 85.8 | #259 (Claude 3.7 Sonnet) |
| 12 | Gemini 3 Pro (Preview) | 85.6 | #71 |
| 13 | GPT-5.1 | 83.9 | #145 |
| 14 | Claude Opus 4.5 | 83.8 | #87 |
| 15 | O3 (2025-04-16) | 83.8 | #125 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=benchtable-stem · How It Works · Data refreshed daily, snapshot 2026-11-05.