OR-License: Claude Code hits 0.87 vs expert 1.0 on real OR work

Claude Code, running as a coding-agent harness, hit 88.39% strict feasibility and a Deployment-Adjusted Quality score of 0.87 against an expert reference of 1 on a new operations-research benchmark built from USD 3.68 million of paid client engagements.

The benchmark, OR-License, comprises 31 instances across 10 industrial scenarios, with original requirements, input data and deliverable formats preserved after de-identification. Across 11 frontier LLMs tested, "several exceed the expert reference in conditional objective quality, but none matches its DAQ," the authors write. DAQ assigns zero to infeasible deliveries, so the aggregate gap is driven by solutions that pass only sometimes.

The paper also introduces ORSim, "a controlled client simulator that discloses business facts without supplying OR expertise," built to study how agents gather what they need from clients. The research chain holds up unevenly: "task-equalized research-chain completion rates range from 24% to 45% across models on evaluable tasks."

The abstract reports a single headline DAQ figure for the strongest primary configuration and does not break out the other ten models individually. The authors frame the findings as motivation for "sustained information-gap resolution, faithful problem formulation, and evidence-linked self-improvement for reliable long-horizon agents."