InvisibleBench: leaderboard
InvisibleBench evaluates caregiver-support AI systems for relational harms, fail-closed safety and compliance gates, communication quality, coordination, and boundary integrity.
Metric: Hard Fail Rate (imported). Source: github.com. Source: benchmarklist.com. Saturation forecast: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 Mini | 0 | #186 |
| 2 | MiniMax-M2.5 | 1.75 | #303 |
| 3 | Claude Sonnet 4.5 | 3.51 | #142 |
| 4 | Grok 4.1 Fast | 3.51 | #217 |
| 5 | GPT-OSS-120B | 5.26 | #341 |
| 6 | Kimi K2.5 | 5.26 | #144 |
| 7 | Qwen 3.5 397B A17B | 5.26 | #146 |
| 8 | GLM-5 | 5.26 | #141 |
| 9 | Gemini 3 Flash | 8.77 | #99 |
| 10 | Gemini 2.5 Flash | 10.53 | #245 |
| 11 | Qwen 3.5 35B A3B | 10.53 | #256 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=invisiblebench · How It Works · Data refreshed daily, snapshot 2026-10-20.