AIBRIEF
Neon and Castform published a reproducible recipe for AI agent multi step lookups (so called agentic search) that, in their own benchmark, matches what the vendors label OpenAI's GPT 5.6 Sol — an unusual model designation — at roughly 1% of the cost.
Neon, the Postgres provider, and Castform, a post-training startup, published a case study claiming a small open-weight model can match GPT-5.6 Sol on multi-hop agentic search at roughly 1% of the cost.
"Agentic search" is the multi-step lookup work that AI agents do across a corpus. Neon and Castform say a typical frontier call on that workload runs more than 10 seconds and costs about $0.03 per request; their post-trained open model, in their own benchmark, finishes the same task at one-hundredth the price.
A GitHub repo and a Neon RAG example describe how Castform generates synthetic queries against Neon's Lakebase Postgres and its new vector-plus-BM25 search extension, runs every reinforcement-learning rollout through that same search tool, then deploys the model calling that same tool at inference. Training and serving share the substrate.
A Hacker News thread on the post is split: some see the pattern as a template for narrow workloads; others argue specialized models rarely beat strong generalists on retrieval and want an independent reproduction.
Two things are not yet settled. The "gpt-5.6-sol" name is unusual and may not be a generally available OpenAI model. The 100x figure is one vendor's self-benchmark on one task, not a production cost claim.