Phare: leaderboard
Giskard multilingual safety probe covering hallucination and reliability, social bias, harmful content, sycophancy, prompt sensitivity, and jailbreak resistance.
Metric: Score (%). Source: phare.giskard.ai. 67 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Haiku 4.5 | 83.16 |
| 2 | Claude Opus 4.5 | 82.38 |
| 3 | Claude Sonnet 5 | 78.1 |
| 4 | Claude Sonnet 4.5 | 77.6 |
| 5 | Claude Opus 4.1 | 76.87 |
| 6 | Kimi K2.6 | 76.06 |
| 7 | Qwen 3.7 Max | 75.57 |
| 8 | Claude Opus 4.6 | 74.13 |
| 9 | Qwen 3.7 Plus | 74.06 |
| 10 | Claude Sonnet 4.6 | 74.03 |
| 11 | Gemini 3 Pro (Preview) | 73.31 |
| 12 | GPT-5 Mini | 73.22 |
| 13 | GPT-5.1 | 72.82 |
| 14 | GPT-5.2 | 71.03 |
| 15 | Llama 4 Maverick | 70.84 |
Interactive version: theaggregate.ai/benchmark?slug=phare · How It Works · Data refreshed daily, snapshot 2026-09-05.