ORAgentBench - Hard: leaderboard
ORAgentBench (2026) tests LLM agents on 107 human-reviewed operations research tasks solved end to end, from multi-file operational data to a validated decision, with hidden validators for feasibility and objective quality; the agents run in the Claude Code or Codex harness. This board is the pass rate (%) on the 34 hard tasks, where a pass needs a feasible submission with normalized quality of at least 0.4; higher is better.
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 June 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 20.59 | #123 (GPT-5.4) |
| 2 | GPT-5.3 Codex (High) | 20.59 | #116 (GPT-5.3 Codex) |
| 3 | Claude Sonnet 4.6 | 14.71 | #161 |
| 4 | Claude Opus 4.6 | 14.71 | #102 |
| 5 | DeepSeek V4 Pro | 13.73 | #188 |
| 6 | Kimi K2.6 | 11.76 | #178 |
| 7 | GLM-5.1 | 8.82 | #179 |
| 8 | DeepSeek V4 Flash | 5.88 | #258 |
| 9 | Qwen 3.6 Plus | 5.88 | #222 |
| 10 | GLM-5 | 5.88 | #260 |
| 11 | GPT-5.4 Mini (High) | 5.88 | #389 (GPT-5.4 Mini) |
| 12 | MiniMax-M2.7 | 2.94 | #518 |
| 13 | Qwen 3.5 Plus | 2.94 | #232 |
| 14 | MiMo-V2.5-Pro | 2.94 | #279 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=oragentbench-hard · How It Works · Data refreshed daily, snapshot 2026-11-13.