Pressure Reveals Character - Robustness: Instruction Hierarchy: leaderboard
Mean 1-5 judge rubric score over 37 multi-turn scenarios testing respecting the precedence of system, developer and user instructions. One behaviour of the Robustness category; a subset of the headline's transcripts.
Metric: Mean judge rubric score over 37 scenarios (1 = failure, 5 = exemplary). Source: storage.googleapis.com. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 4.86 | #741 |
| 2 | GPT-5 Mini | 4.59 | #348 |
| 3 | GPT-OSS-120B | 4.38 | #700 |
| 4 | Claude Sonnet 4.5 | 4.16 | #250 |
| 5 | Llama 4 Maverick | 4.14 | #1099 |
| 6 | Llama 3.1 405B Instruct | 3.97 | #1115 |
| 7 | Claude Sonnet 4 | 3.95 | #400 |
| 8 | Claude Opus 4.5 | 3.92 | #156 |
| 9 | Gemini 3 Pro | 3.89 | #137 |
| 10 | Mistral Medium 3.1 | 3.89 | #852 |
| 11 | Qwen 3 235B A22B | 3.86 | #623 |
| 12 | Grok 4.1 | 3.86 | #476 |
| 13 | Gemini 3 Flash | 3.84 | #164 |
| 14 | Llama 3.3 70B | 3.84 | #1498 |
| 15 | Gemini 2.5 Pro | 3.68 | #266 |
Interactive version: theaggregate.ai/benchmark?slug=pressure-reveals-character-robustness-instruction-hierarchy · How It Works · Data refreshed daily, snapshot 2026-11-13.