Castform + Neon trounce GPT-5.6 Sol on retrieval tasks
A 4B open model trained with Castform matches GPT-5.6 Sol's search accuracy for 100x less cost.
Castform and Neon have co-developed a breakthrough approach to agentic search that undercuts the cost and latency of frontier models like GPT-5.6 Sol by a factor of 100. Their solution leverages a lightweight 4B open-source model, RL post-trained on proprietary enterprise data stored in Neon's Lakebase Postgres—a vectorized Postgres extension. The model matches the retrieval accuracy of leading closed models but achieves this at a fraction of the cost and latency, making multi-turn agentic search practical for mainstream adoption.
The core innovation lies in Castform's ability to transform raw enterprise data (e.g., internal documentation, customer interactions, operational databases) into synthetic training tasks and reward functions without requiring extensive data engineering. By integrating with Neon's Lakebase Search, Castform automates the RL post-training loop, enabling developers to deploy production-grade agentic search systems without managing GPU infrastructure or complex ML pipelines. This democratizes access to high-performance, cost-efficient AI agents for retrieval tasks.
- Castform + Neon's 4B open model achieves GPT-5.6 Sol-level retrieval accuracy at 100x lower cost.
- RL post-training is automated, turning raw enterprise data into synthetic tasks without manual labeling.
- Neon's Lakebase Search (Postgres extension) powers the retrieval environment for training and inference.
Why It Matters
Unlocks scalable, affordable agentic search for enterprises by replacing costly frontier models with cost-efficient open alternatives.