OptiVerse - Hard: leaderboard
Metric: Accuracy (%) on the 300 hard-level problems of OptiVerse; the model writes and runs solver code (gurobi, pyomo, cvxpy, ortools and others) and a problem counts as solved only when every required variable and the objective match the ground truth within 0.1% relative error, extracted and checked by a DeepSeek-V3.2-Chat judge; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 22 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 27 |
| 2 | Gemini 3 Flash | 25.33 |
| 3 | GPT-5.2 (Thinking) | 25.33 |
| 4 | Claude Sonnet 4.5 (Thinking) | 21.67 |
| 5 | DeepSeek V3.2 (Thinking) | 21.33 |
| 6 | O3 | 20.67 |
| 7 | O4 Mini | 20.67 |
| 8 | Gemini 2.5 Pro | 20 |
| 9 | DeepSeek V3.2 (Non-reasoning) | 19 |
| 10 | Gemini 2.5 Flash | 18.67 |
| 11 | GPT-OSS-120B | 18.67 |
| 12 | Kimi K2 | 16.33 |
| 13 | Qwen 3 8B (Thinking) | 16 |
| 14 | Qwen 3 Coder 30B A3B Instruct | 11.67 |
| 15 | Qwen 3 8B (Non-reasoning) | 7.67 |
Interactive version: theaggregate.ai/benchmark?slug=optiverse-hard · How It Works · Data refreshed daily, snapshot 2026-10-07.