Harbor-Index 1.0 (Terminus-2): leaderboard
Harbor-Index 1.0 (the Harbor framework team with Stanford, Cornell Tech, Peking University and others, 2026) is 82 hard agentic tasks curated from 29 benchmarks ported to the Harbor framework, spanning software engineering, scientific research, tool use, knowledge, mathematics, data analytics and security, each kept only after a difficulty filter and AI and human audits for broken or exploitable tasks. The score is the pass rate in percent over the 82 tasks, one run per task with default model settings, higher is better; continuous tasks such as GSO and AlgoTune speedups pass only at a locked threshold set by the best observed result. On this board every model runs in the Terminus-2 harness, and tasks were kept only where Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro together solved at most a third of 18 trials, so the item set is not independent of Gemini 3.1 Pro.
Metric: Pass rate (%; 82 tasks, one run each). Source: arxiv.org. Saturation forecast: Around October 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 19.5 | #48 |
| 2 | Claude Opus 4.8 | 15.9 | #56 |
| 3 | Gemini 3.1 Pro (Preview) | 11 | #91 |
| 4 | GLM-5.2 | 9.8 | #137 |
| 5 | Kimi K2.6 | 8.5 | #173 |
| 6 | MiniMax-M3 | 6.1 | #265 |
| 7 | Qwen 3.7 Max | 4.9 | #142 |
| 8 | DeepSeek V4 Pro | 3.7 | #181 |
| 9 | MiMo-V2.5-Pro | 2.4 | #274 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=harbor-index-1-0-terminus-2 · How It Works · Data refreshed daily, snapshot 2026-11-12.