Post-trained open models can undercut frontier LLMs at agentic search by 100x

Retrieval has moved from one-shot RAG to agentic search, where a model plans and queries in a loop, hitting a frontier LLM on every iteration. That pattern gets expensive fast: the post cites a typical multi-turn search against gpt-5.6-sol running over 10 seconds and roughly $0.03 per request. Small open-weights models are about 100x cheaper but weaker out of the box, and the pitch here is that reinforcement-learning post-training closes that gap — on a narrow task like search, a tuned open model can match or beat the frontier at a fraction of the cost and latency.

The piece is a joint vendor pitch from Neon and a startup called Castform. Castform’s angle is that the training data most teams need is already sitting in their corpus — internal docs, support articles, wikis, product records — and the real barriers are converting that into labeled tasks and running the RL loop without an ML team. It claims to automate both: synthesizing question/ground-truth pairs from existing documents, letting developers specify the agent’s tools and a reward function (retrieve the right chunks, cite sources, answer correctly), then managing the trial-and-error loop with observability to catch broken tools or reward hacking.

Neon’s role is the retrieval and environment layer. Its Lakebase Postgres and new Search extensions serve queries during training, and its autoscaling is meant to absorb the bursty load of thousands of parallel rollouts without paying for peak capacity around the clock. More interesting for stateful agents that mutate data: Neon’s branching gives each rollout an isolated, resettable database, and time-travel queries let you reconstruct the state an agent saw. Worth reading as a marketing argument rather than an independent benchmark — the headline numbers and the ‘beats frontier models’ framing come from the vendors themselves — but the underlying trend it points at, task-specific RL tuning of cheap open models to displace general-purpose frontier calls, is a real cost lever for teams running agentic retrieval at scale.