Neon AI Model Matches GPT-5.6 Sol at 100x Lower Cost

A 4-billion-parameter open model reportedly matched GPT-5.6 Sol on a specific search-and-retrieval task this week. This model achieved its performance by being trained exceptionally well for one job. The claim comes from a blog post by Neon, a database company, and Castform, a startup specializing in fine-tuning open models. They state their model matched GPT-5.6 Sol's retrieval accuracy while costing about 100 times less per request. Instead of using a large, general model, they trained a smaller one specifically for retrieval using reinforcement learning. A typical multi-turn search request handled by GPT-5.6 Sol can take over 10 seconds and cost around $0.03. The specialized model repeatedly practiced retrieving correct information and optimizing for that single workflow. This highlights a shift in AI development, where specialized models can compete with larger ones on specific tasks without needing broad general intelligence. This trend could lead to more efficient and cost-effective AI solutions for focused tasks.