Neon and Castform have described a retrieval experiment in which a post-trained, four-billion-parameter open model matched the search accuracy of a frontier model while operating at a claimed fraction of the cost. The result concerns a specific agentic retrieval setup, not a general comparison of model intelligence.
The project combines Castform's reinforcement-learning post-training pipeline with search over data held in Neon's Postgres platform. Its premise is that companies already possess useful training material in knowledge bases and operational databases, but converting that material into tasks, correct answers and rewards normally requires extensive preparation. Castform generates question-and-answer tasks from the corpus and manages repeated training attempts.
In each attempt, an agent can call Lakebase Search until it has enough material to answer. A reward function scores whether the system retrieved the correct passages, cited the intended sources and returned the right response. Feedback from those scores then guides later attempts. The developers say observability tools let users inspect both the aggregate reward trend and individual runs for problems such as broken tools or reward hacking.
The companies frame the work around a shift from one-shot similarity search toward multi-step retrieval. An agent may break a question into smaller searches and make several model calls before composing an answer. That can improve the process, but repeated use of an expensive model increases both latency and cost. According to the project account, a representative multi-turn request using the compared frontier system took more than 10 seconds and cost about three cents end to end.
The reported 100-fold cost difference reflects the economics of running the smaller open model for this focused task. It should not be interpreted as proof that a 4B model replaces a frontier system across unrelated workloads. The published evidence describes an internally constructed training and evaluation pipeline and does not provide an independent replication in the supplied material.
Neon says its infrastructure can scale for bursts created by thousands of parallel training rollouts and reduce compute during idle periods. It also points to database branching as a way to isolate future agents that modify state, preventing one training attempt from affecting another or a production database.
For organizations, that changes the optimization question from selecting one model for everything to deciding where specialization justifies training and operational complexity.
The demonstration presents post-training as an alternative to paying for a general-purpose frontier model at every retrieval step. Its practical appeal will depend on whether teams can define reliable tasks and reward signals, and whether the savings persist when their own data, search patterns and quality requirements replace the example environment.
