Beating GPT-5.6 Sol on retrieval with 100x cheaper open models

**Castform** and **Neon** reported that a **4B open-source model** post-trained with Castform retrieved search results as accurately as **GPT-5.6 Sol** while costing about **100x less**. The write-up, credited to Pranav Aurora, Ying Hang Seah, and Angel Pan and dated Aug 05, 2026, frames the result as a price-and-efficiency win over frontier models rather than a raw capability leap.

On the product side, the stack splits agentic search into two jobs. **Neon** (Lakebase Postgres) and its new **Search extensions** handle context: tools that let an agent find the right data. **Castform** handles the model side: teaching the model what to search for. Ying Hang Seah, cofounder of Castform, argues that most teams already hold their best training data in databases, but turning that data into something usable is hard, and cheap, large-scale agent read/search/mutate work needs serious infrastructure—pointing Castform at Neon is meant to skip both steps.

For builders, the practical claim is that retrieval quality need not track frontier API spend. A post-trained **4B** open model matching **GPT-5.6 Sol** on search accuracy at roughly **1/100th** the cost shifts the bottleneck from model size to data access and post-training. Teams with production Postgres data, especially on Neon, can treat their own tables as the training and retrieval surface instead of shipping everything through a large closed model for every lookup.

That sits against a market that, around **2022**, went all-in on embedding search as the default retrieval path. Castform’s pitch is that the hard part is no longer only similarity over vectors, but whether the model can decide *what* to search when tools and live database context are available. Neon’s search extensions and Castform’s post-training sit on opposite ends of that loop: infra for finding data, model training for choosing the right queries.

Watch whether other teams reproduce the same accuracy-vs-cost gap with open **4B**-class models on real workloads, and how far Neon’s search extensions plus Castform-style post-training generalize beyond the reported retrieval setup. The concrete next check for engineers is whether pointing a small post-trained model at their own Neon data closes the gap with frontier retrieval without rewriting the agent stack.