RealGuardrails: leaderboard
The model gets one of 14 realistic assistant system prompts with guardrails and a hand-written user message that complies with or conflicts with them; the number is the % of 239 cases judged as following the system prompt. Unlike prompt-injection tests, the user asks plainly, without adversarial tricks.
Metric: Handwritten guardrail test pass rate (%, higher is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 Mini | 83.3 | #533 |
| 2 | Gemini 1.5 Flash (002) | 65.7 | #1464 |
| 3 | GPT-4o (2024-08-06) | 65.3 | #740 |
| 4 | GPT-4o Mini (2024-07-18) | 64.4 | #1510 |
| 5 | Gemini 1.5 Flash-8B (001) | 59 | #2331 |
| 6 | DeepSeek V3 | 50.2 | #716 |
| 7 | Llama 3.1 8B Instruct | 47.3 | #4246 |
| 8 | DeepSeek R1 | 46.9 | #473 |
| 9 | Llama 3 8B Instruct | 41.8 | #4612 |
| 10 | Llama 3.2 3B Instruct | 37.7 | #7147 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=realguardrails · How It Works · Data refreshed daily, snapshot 2026-11-13.