Neon and Castform: a Post-Trained 4B Open Model Matches GPT-5.6 Sol on Agentic Retrieval at 1/100th the Cost
Neon (corroborated by Hacker News front page, 341 points)·high signal
Neon published on August 5 that a 4B open-source model post-trained with Castform retrieved search results as accurately as GPT-5.6 Sol while costing 100x less. Their baseline: a typical multi-turn agentic search request against gpt-5.6-sol takes more than 10 seconds and roughly $0.03 end-to-end, because retrieval has shifted from one-shot search to models planning and searching in a loop — which multiplies frontier token cost by the number of hops. For anyone building RAG or agentic search, this reframes the frontier model as the teacher rather than the runtime: post-train a small open model on your retrieval loop and the per-request economics change by two orders of magnitude.