Planning-only workers reportedly matched full-stage coding agents' reward at half the cost

A user describing a Cohere paper says researchers tested agents on ProgramBench, where they rebuild programs from executables and documentation. Planning-only workers matched the reward of workers used at every stage at half the cost. A GPT-5.6 Sol lead with 16 cheaper Luna planners came within 1.4 points of an all-Sol team at 70% lower cost; cheaper agents used for coding lowered reward.