BoundaryBench (Non-root): leaderboard
BoundaryBench success on all 89 Terminal-Bench 2.1 tasks, three trials per model-harness bundle at high reasoning effort, with the coding agent running as an ordinary non-root user and everything else open. Permission Denied Appendix Table 12 reports twelve bundles; percent of passing trials, higher is better. The paper reports the Claude Fable 5 gated route with its declared Opus 4.8 fallback: across all 801 trials over three policies, 27.6% of trials touched the fallback and 29.9% of requests used it; these pooled shares are not policy-specific.
Metric: Success Rate under non-root policy (%). Source: arxiv.org. Saturation forecast: Around September 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.5 (High) | 73 | #54 (Grok 4.5) |
| 2 | GPT-5.6 Sol (High) | 72.3 | #22 (GPT-5.6 Sol) |
| 3 | Claude Opus 5 (High) | 68.2 | #16 (Claude Opus 5) |
| 4 | Claude Fable 5 (High) | 65.9 | #20 (Claude Fable 5) |
| 5 | GPT-5.6 Terra (High) | 65.2 | #40 (GPT-5.6 Terra) |
| 6 | Claude Opus 4.8 (High) | 62.9 | #39 (Claude Opus 4.8) |
| 7 | Claude Sonnet 5 (High) | 60.3 | #65 (Claude Sonnet 5) |
| 8 | Kimi K3 (High) | 60.3 | #35 (Kimi K3) |
| 9 | GLM-5.2 (High) | 59.6 | #80 (GLM-5.2) |
| 10 | GPT-5.6 Luna (High) | 56.6 | #79 (GPT-5.6 Luna) |
| 11 | MiniMax-M3 (High) | 46.8 | #151 (MiniMax-M3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=boundarybench-non-root · How It Works · Data refreshed daily, snapshot 2026-11-05.