Neon and Castform Launch Efficient Search Model at 1/100th the Cost of GPT-5.6 Sol

Neon, a serverless PostgreSQL database provider, has teamed up with Castform, an AI voice agent platform, to launch a training method for search models. Using reinforcement learning, they fine-tuned a 4-billion-parameter open-source model to help AI agents handle complex searches more efficiently.

In document search, traditional embedded search is giving way to agentic search, where models break down queries and iterate on results. This usually requires powerful models, leading to high costs and delays. Neon and Castform's approach trains models to find the right information faster, cutting costs dramatically.

In Neon's tests, the fine-tuned model scored 1.447, beating GPT-5.6 Sol's 1.369 and GPT-5.4's 1.377. Its per-query cost is only $0.000929, compared to GPT-5.6 Sol's $0.087338—about 1/100th the price.