Neon and Castform's 4B open-source model matches GPT-5.6 Sol retrieval at 100x lower cost

A 4-billion-parameter open-source retrieval model matched OpenAI’s GPT-5.6 Sol on multi-turn agentic search accuracy at roughly 100 times lower cost, Postgres platform Neon and post-training startup Castform said on August 5, 2026.

Retrieval models fetch the right documents for an AI system to reason over, and their cost compounds across the many search steps an agent takes. In a joint post on Neon’s blog, the companies said the model was reinforcement-learning post-trained by Castform using Neon’s Lakebase Search, a Postgres-based search extension, so the training data came straight from a live database rather than a hand-built corpus.

By the companies’ measurement, GPT-5.6 Sol took more than 10 seconds and cost about $0.03 per end-to-end multi-turn search request; the post-trained 4-billion-parameter model matched its accuracy at roughly 100 times lower cost. Castform cofounder Ying Hang Seah said most teams’ best training data “is just sitting in their databases” and is hard to turn into usable training data.

The figures come from a single vendor blog post and have not been independently verified. The companies did not publish a full benchmark breakdown, and “agentic search accuracy” can vary widely by task, so the 100x cost gap should be read as their own claim rather than a settled result.

If the approach holds up outside the vendors’ own tests, cheap task-specific retrieval models could pull agent workloads away from frontier APIs. The open question is whether independent teams reproduce the cost gap on their own data.