OR-Space - Explain: leaderboard

The Explain task of OR-Space (2026), a workspace benchmark for industrial optimization agents: from the workspace and solver outputs such as logs, duals and slacks, the agent writes short factual reports on bottlenecks and sensitivities. Scored 0-100 by a rubric with a hallucination penalty; higher is better.

Metric: Explain rubric score (0-100): exact coverage, reasoning, grounding and answer quality of short solver-grounded reports minus a hallucination penalty, checklist items scored programmatically or by a judge model, 100 industrial optimization instances in a filesystem workspace (docs, data and an empty or heuristic src directory), Gurobi 12.0.1 track, API temperature at most 0.1; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.486.52#135
2GPT-5 Mini82.94#364
3Claude Opus 4.680.2#113
4Claude Sonnet 4.5Thinking79.63#261
5Qwen 3.5 27B77.88#375
6DeepSeek V4 Flash77.77#267
7GPT-5.176.16#240
8Qwen 3 Max74.66#415
9Gemini 3.1 Pro (Preview)73#99
10DeepSeek V4 Pro68.87#196
11Qwen 3 32BNon-reasoning61.32#972
12Qwen 3 14BNon-reasoning61.03#1269
13GPT-4o59.36#751
14Qwen 3 8BNon-reasoning55.19#1873
15Gemini 3 Flash53.24#175

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=or-space-explain · How It Works · Data refreshed daily, snapshot 2026-11-21.