ORAgentBench - Hard: leaderboard

Metric: Pass rate (%) on the 34 hard tasks: a task passes when the submission is feasible and its normalized quality reaches 0.4; end-to-end operations research tasks packaged with multi-file operational data, configuration and a submission schema; the agent writes and runs solver code in an isolated Harbor environment (45-minute interaction budget, 5-minute solve limit) and hidden validators score schema and hard-constraint feasibility and normalized objective quality; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 14 models tracked.

Top models

#ModelScore
1GPT-5.4 (High)20.59
2GPT-5.3 Codex (High)20.59
3Claude Sonnet 4.614.71
4Claude Opus 4.614.71
5DeepSeek V4 Pro13.73
6Kimi K2.611.76
7GLM-5.18.82
8DeepSeek V4 Flash5.88
9Qwen 3.6 Plus5.88
10GLM-55.88
11GPT-5.4 Mini (High)5.88
12MiniMax-M2.72.94
13MiMo-V2.5-Pro2.94
14Qwen 3.5 Plus2.94

Interactive version: theaggregate.ai/benchmark?slug=oragentbench-hard · How It Works · Data refreshed daily, snapshot 2026-09-29.