BLINDSPOT: leaderboard
Safety Calibration Score for long-horizon tool-using agents under adversarial scenarios: a composite that rewards refusing unsafe requests while completing benign ones. Higher is better. Source: arxiv.org/abs/2609.16305.
Metric: Safety Calibration Score. Source: arxiv.org. Saturation forecast: Around August 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 96 | #22 |
| 2 | Llama 3.3 70B | 95 | #578 |
| 3 | GPT-5.6 Luna | 93 | #79 |
| 4 | GPT-5.6 Terra | 90 | #40 |
| 5 | Gemini 2.5 Pro | 88.5 | #159 |
| 6 | GPT-4.5 | 85.5 | #278 |
| 7 | Claude Opus 4.6 | 83.1 | #67 |
| 8 | GPT-4o | 80.9 | #357 |
| 9 | Gemma 3 27B | 76.3 | #606 |
| 10 | Gemma 3 12B | 69.5 | #674 |
| 11 | Claude Haiku 4.5 | 63.8 | #287 |
| 12 | GPT-3.5 Turbo | 62.6 | #858 |
| 13 | Mistral Large 3 | 50.6 | #410 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=blindspot · How It Works · Data refreshed daily, snapshot 2026-11-05.