StartupBench: leaderboard
Importance-weighted mean task score from 0 to 100 on 97 workflow tasks, averaged over three runs under the Nanobot harness with a 200-step cap. This is distinct from the fraction of tasks scoring at least 90.
Metric: Mean Task Score (0-100, average of three runs). Source: arxiv.org. Saturation forecast: Around January 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 73.67 | #30 |
| 2 | GPT-5.6 Sol | 73.61 | #19 |
| 3 | GPT-5.5 | 72.79 | #31 |
| 4 | Seed 2.1 Pro | 67.19 | #67 |
| 5 | DeepSeek V4 Pro | 61.11 | #102 |
| 6 | GLM-5.1 | 60.79 | #97 |
| 7 | Kimi K2.6 | 59.95 | #104 |
| 8 | Qwen 3.6 Max | 59.46 | #96 |
| 9 | Gemini 3.1 Pro (Preview) | 49.73 | #58 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=startupbench · How It Works · Data refreshed daily, snapshot 2026-10-20.