Castform + Neon Enable 4B Models to Match GPT-5.6 Sol at 100x Lower Cost
Key Takeaways
- ▸A 4B open-source model post-trained with Castform matches GPT-5.6 Sol's retrieval accuracy at 100x lower cost, shifting the economic calculus for enterprise AI deployments
- ▸Agentic retrieval workflows (multi-hop, iterative search) have exposed cost and latency problems with frontier models; smaller post-trained models now offer a viable alternative
- ▸Castform democratizes RL post-training by turning unstructured enterprise data into training tasks, eliminating the need for ML teams and GPU infrastructure
Summary
Castform has demonstrated that a 4B open-source model post-trained with their RL platform can retrieve search results as accurately as OpenAI's GPT-5.6 Sol while costing 100 times less. The company has partnered with Neon (Lakebase Postgres) to enable this capability, leveraging existing enterprise data stored in databases as training material.
The breakthrough addresses a critical inefficiency in agentic AI systems. As multi-turn search workflows have become standard (moving beyond single-query RAG), frontier models like GPT-5.6 Sol incur significant costs (~$0.03 per request) and latency (>10 seconds), making them economically prohibitive at scale. Meanwhile, small open-source models lack the retrieval capabilities to replace them—until now.
Castform's approach turns this on its head by enabling organizations to RL post-train open-source models directly on their proprietary knowledge bases—internal documentation, customer interactions, and operational databases. The platform abstracts away ML infrastructure complexity, making advanced model customization as accessible as prompt engineering. Neon's infrastructure handles the data retrieval and tooling layer that agents need during training.
- The Castform + Neon partnership combines model post-training with database-native search, enabling organizations to build agents that can efficiently use their existing data
Editorial Opinion
This is a significant step toward making advanced AI capabilities accessible to organizations of all sizes. When open-source models can match—or exceed—frontier model performance at 1/100th the cost on high-value tasks like retrieval, it fundamentally changes the business case for AI adoption. However, it's crucial to note this is specialized to retrieval and search; whether similar breakthroughs generalize to reasoning, planning, and other complex tasks remains the open question that will determine whether open models truly displace closed APIs.
