HarnessAudit-Bench - Task Completion: leaderboard
Checkpointed task completion on HarnessAudit-Bench's 210 benign multi-agent workflow tasks (Liu et al., 2026): weighted completion checkpoints verified from the trajectory, backend state or final output, by rules and a GPT-5.4 judge, 0-1, higher is better. It measures completion alone; the paper's finding is that completion and safe execution diverge.
Metric: Task completion rate (0-1; weighted hidden completion checkpoints, rule-based and GPT-5.4-judged, averaged over 210 tasks). Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 0.74 | #100 |
| 2 | Qwen 3.5 Plus | 0.69 | #227 |
| 3 | GLM-5V Turbo | 0.68 | #342 |
| 4 | GPT-5.4 | 0.66 | #121 |
| 5 | Claude Sonnet 4.6 | 0.64 | #158 |
| 6 | Gemini 3.1 Pro (Preview) | 0.56 | #91 |
| 7 | Kimi K2.6 | 0.54 | #173 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harnessaudit-bench-task-completion · How It Works · Data refreshed daily, snapshot 2026-11-12.