SafePyramid - Exact Match: leaderboard
Exact-match rate on SafePyramid (2026), an in-context policy guardrailing benchmark: the percent of the 3,000 cases for which the predicted set of violated rules equals the ground-truth set, with refused cases excluded. Higher is better.
Metric: Rule matching rate (%) over all 3,000 SafePyramid cases: per case (a multi-turn conversation plus an in-context safety policy), the predicted set of violated rules matches the ground-truth set when neither false positives nor false negatives exceed (1 - tau) of the true violations; at tau = 1.0 only, i.e. the exact violated-rule set; per-policy evaluation, refused cases excluded; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (Max) | 35 | #79 (Claude Opus 4.7) |
| 2 | GPT-5.5 (xHigh) | 34.9 | #49 (GPT-5.5) |
| 3 | DeepSeek V4 Pro (Max) | 31.7 | #188 (DeepSeek V4 Pro) |
| 4 | Seed 2.0 Pro (High) | 29.5 | #211 (Seed 2.0 Pro) |
| 5 | Qwen 3.6 Max (Preview) (Thinking) | 29.2 | |
| 6 | Gemini 3.5 Flash (High) | 28.1 | #101 (Gemini 3.5 Flash) |
| 7 | Hy3-preview (High) | 27.4 | #316 (Hy3-preview) |
| 8 | Grok 4.3 (High) | 25.1 | #333 (Grok 4.3) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=safepyramid-exact-match · How It Works · Data refreshed daily, snapshot 2026-11-13.