ExCyTIn-Bench: leaderboard
ExCyTIn-Bench (Microsoft) has LLM agents investigate eight multi-stage attacks by querying a MySQL copy of Microsoft Sentinel logs from a controlled Azure tenant, answering 589 test questions within 25 steps at temperature 0. Reward is 1 for a correct answer, else decayed partial credit for intermediate evidence, graded by GPT-4o; the board shows the total average from 0 to 1, higher is better.
Metric: Total average reward over the 589 test questions (0-1, partial credit). Source: github.com. Saturation forecast: Around February 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 | 0.61 | #87 |
| 2 | GPT-5.1 (High) | 0.58 | #145 (GPT-5.1) |
| 3 | GPT-5 (High) | 0.56 | #105 (GPT-5) |
| 4 | GPT-5 (Medium) | 0.54 | #105 (GPT-5) |
| 5 | GPT-5.1 (Medium) | 0.54 | #145 (GPT-5.1) |
| 6 | Claude Sonnet 4.5 | 0.49 | #148 |
| 7 | O3 | 0.46 | #132 |
| 8 | GPT-5.1 (Low) | 0.45 | #145 (GPT-5.1) |
| 9 | GPT-5 (Low) | 0.38 | #105 (GPT-5) |
| 10 | GPT-5 Mini | 0.37 | #193 |
| 11 | O4 Mini | 0.37 | #185 |
| 12 | Grok 4 | 0.34 | #189 |
| 13 | GPT-4.1 | 0.34 | #258 |
| 14 | GPT-5.1 (Non-reasoning) | 0.32 | #145 (GPT-5.1) |
| 15 | Gemini 2.5 Flash | 0.31 | #255 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=excytin-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.