IH-Benchmark: leaderboard
Scenario-weighted instruction-hierarchy compliance over both conflict and non-conflict scenarios and both system-over-user and user-over-tool tracks, Table 15. Model reasoning variants remain separate; this is not the conflict-only headline table.
Metric: Overall Compliance (%, conflict and non-conflict). Source: arxiv.org. Saturation forecast: Estimated already saturated. 37 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 98.3 | #67 |
| 2 | GPT-5.4 (Medium) | 97.6 | #83 (GPT-5.4) |
| 3 | Claude Opus 4.5 | 97.3 | #87 |
| 4 | Claude Opus 4.7 | 96.7 | #51 |
| 5 | Claude Sonnet 4.6 | 96.1 | #94 |
| 6 | GPT-5.4 (Low) | 94.1 | #83 (GPT-5.4) |
| 7 | Grok 4.1 Fast (Reasoning) | 89.9 | #227 (Grok 4.1 Fast) |
| 8 | GPT-5.2 (Medium) | 89.3 | #115 (GPT-5.2) |
| 9 | GPT-5 Mini (Medium) | 88.6 | #193 (GPT-5 Mini) |
| 10 | Claude Haiku 4.5 | 88.3 | #287 |
| 11 | GPT-5.2 (Low) | 87.8 | #115 (GPT-5.2) |
| 12 | Claude Sonnet 4.5 | 87.6 | #148 |
| 13 | GLM-5 | 86.6 | #149 |
| 14 | GLM-5.1 | 86 | #109 |
| 15 | GPT-5 Mini (Low) | 85.8 | #193 (GPT-5 Mini) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ih-benchmark · How It Works · Data refreshed daily, snapshot 2026-11-05.