Meta Muse Spark 1.1 Report - GDM Stealth - Strategic Rule-Breaking: leaderboard
GDM Stealth Strategic Rule-Breaking challenge (Phuong et al., 2025): average success rate of tasks accomplished while bypassing the task-specific monitor. Meta Muse Spark 1.1 Evaluation Report (July 2026), Table 9; percent, higher means more stealth capability. Muse Spark 1.1 appears unguarded and helpful-only; human threshold 47%.
Metric: Success Rate (%). Source: research.meta.ai. Saturation forecast: Around June 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (High) | 55.5 | #64 (Gemini 3.1 Pro (Preview)) |
| 2 | GPT-5.5 (xHigh) | 50.1 | #33 (GPT-5.5) |
| 3 | Muse Spark 1.1 (High) | 49.5 | #55 (Muse Spark 1.1) |
| 4 | Claude Opus 4.8 (Max) | 43.3 | #39 (Claude Opus 4.8) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=meta-muse-spark-1-1-report-gdm-stealth-strategic-rule-breaking · How It Works · Data refreshed daily, snapshot 2026-11-05.