AlgoWorlds - Reference Utility: leaderboard
AlgoWorlds (2026) turns combinatorial optimization problems from ten families (such as routing, siting, dispatch and portfolio selection) into 240 partially observed environments where an agent gathers information through tools and then commits one decision. Reference utility scores a feasible decision from 0 at a predefined suboptimal reference to 1 at the global optimum (infeasible decisions 0), shown in percent over three trials; higher is better.
Metric: Reference utility (%; x100 of a 0-1 score per feasible decision, 1 at the global optimum and 0 at a predefined suboptimal reference decision, 0 when infeasible; 240 partially observed environments built from ten combinatorial optimization families at four workload levels; the agent sees a hidden instance only through task-specific information tools, then commits one structured decision checked by an independent verifier; every model at its maximum reasoning setting; mean of three trials). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 74.98 | #29 (GPT-5.6 Sol) |
| 2 | Claude Opus 4.8 (Max) | 73.14 | #56 (Claude Opus 4.8) |
| 3 | Claude Sonnet 5 (Max) | 67.2 | #108 (Claude Sonnet 5) |
| 4 | GPT-5.6 Terra (Max) | 54.1 | #60 (GPT-5.6 Terra) |
| 5 | GLM-5.2 (Max) | 35.48 | #137 (GLM-5.2) |
| 6 | Qwen 3.5 Plus | 15.82 | #227 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=algoworlds-reference-utility · How It Works · Data refreshed daily, snapshot 2026-11-12.