Open 4B Model Matches GPT-5.6 Sol at 100x Lower Cost
Original: Beating GPT-5.6 Sol on retrieval with 100x cheaper open models
Why This Matters
RL post-training on small open models could significantly reduce costs for production-scale agentic AI deployments.
Castform and Neon announced that a 4-billion-parameter open-source model, post-trained via reinforcement learning, matches GPT-5.6 Sol on agentic retrieval accuracy while costing approximately 100x less per request, using Neon's Lakebase Postgres and Search extensions as the data backbone.
Castform and Neon published details on how a 4B open-source model trained with reinforcement learning (RL) post-training achieves retrieval accuracy comparable to GPT-5.6 Sol. A typical multi-turn agentic search request using GPT-5.6 Sol takes over 10 seconds and costs roughly $0.03 end-to-end, making it slow and expensive at scale. Small open-weight models cost about 100x less but historically lagged behind closed API models in capability. Castform's RL post-training pipeline bridges this gap for specific tasks like search, without requiring teams to manage ML or GPU infrastructure—aiming to make post-training as accessible as prompt engineering. The pipeline relies on Neon's Lakebase Postgres for corpus storage, and uses the lakebase_text and lakebase_vector extensions for synthetic training data generation. Castform co-founder Ying Hang Seah noted: 'Most teams' best training data is just sitting in their databases. Pointing Castform at Neon skips both [data prep and infra challenges].' The approach reflects a broader industry shift from one-shot RAG pipelines toward multi-hop agentic retrieval, where models plan and issue multiple search queries in a loop.