AuthorizationBench - Claude SDK - Safe Completion Rate: leaderboard
AuthorizationBench's safe completion rate for Claude Agent SDK 0.2.159 on 286 executable business tasks. Each model has separate low/high reasoning runs under a 300-second, 35-turn budget; an in-flight turn may finish later. Safe completion requires finishing without acting outside user authorization. All assigned trajectories remain in the denominator; MiMo v2.6 Pro supplies oracle-user replies and Claude Opus 5.5 judges outcomes. Source API/harness/effort identities are preserved.
Metric: Safe Completion Rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (High) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 79.4 | |
| 2 | Claude Sonnet 5 (Low) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 78.7 | |
| 3 | Claude Sonnet 5 (High) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 78 | |
| 4 | GLM-5.3 (High) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 78 | |
| 5 | GPT-5.6 Sol (Low) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 77.3 | |
| 6 | Gemini 3.8 Flash (Low) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 74.1 | |
| 7 | GLM-5.3 (Low) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 71.7 | |
| 8 | Gemini 3.8 Flash (High) (AuthorizationBench Run: Claude Agent SDK 0.2.159; 300s; 35 turns) | 65 |
Interactive version: theaggregate.ai/benchmark?slug=authorizationbench-claude-sdk-safe-completion-rate · How It Works · Data refreshed daily, snapshot 2026-11-05.