Open-source models beat GPT-4.5 on retrieval tasks at 100x lower cost

A team at Neon has demonstrated that open-source language models can outperform frontier models like GPT-4.5 on retrieval benchmarks while costing roughly 100 times less. The approach, developed in collaboration with Castform, focuses on optimizing retrieval-augmented generation pipelines using smaller, efficient models. The findings were published on Neon's engineering blog and have drawn attention in developer communities. The work highlights a growing trend of cost-effective open models closing the performance gap with expensive proprietary systems.