Beating GPT-5.6 Sol on retrieval with 100x cheaper open models

A 4B open-source model trained using Castform's reinforcement learning post-training matched GPT-5.6 Sol's accuracy on retrieval tasks while costing 100x less, demonstrating that small models can compete with frontier models on specific tasks when properly optimized. The approach shifts from traditional embedding-based RAG to agentic retrieval, where models perform multiple search iterations in a loop, with Castform and Neon's infrastructure enabling this training without requiring deep ML expertise. This makes advanced AI capabilities accessible and cost-effective for developers by leveraging existing database data as training material.