DeliverBench - Final delivery - P01: leaderboard
DeliverBench Table 5:11 persistent-workspace professional projects,150 sequential subtasks,seven exact model/harness systems. Harbor 0.20.0;4-hour task limit,1200-second setup; versions ClaudeCode 2.1.278,Codex 0.142.4,QwenCode 0.24.1,KimiCLI 1.50,GeminiCLI 0.60.0. Launcher requests high reasoning where supported; unknown per-provider support is not invented. GPT-5.6 Luna high judge applies expert signed rubrics; final delivery retains fatal acceptance gates, process averages submitted checkpoints. Native 0–100 rubric scores, not success rates; projects equally weighted in Mean. Metric: Rubric score (0–100).
Metric: Rubric score (0–100). Source: arxiv.org. Saturation forecast: Around May 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 58.84 | #35 |
| 2 | Claude Opus 5 | 58.6 | #16 |
| 3 | GPT-5.6 Sol (High) | 54 | #22 (GPT-5.6 Sol) |
| 4 | GLM-5.2 | 50.36 | #80 |
| 5 | Qwen 3.8 Max | 47.94 | #34 |
| 6 | DeepSeek V4.1 Flash | 47.22 | #37 |
| 7 | Gemini 3.8 Flash | 26.63 | #30 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=deliverbench-final-delivery-p01 · How It Works · Data refreshed daily, snapshot 2026-11-05.