- Published on
Document Search Surpasses GPT-5.6 Sol, at 1/100th the Cost
- Authors
- Name
- DeepHorus
Neon has recently partnered with Castform to release a training solution for retrieval models and an open-source model with 4 billion parameters. In document search tasks, this model achieves accuracy comparable to or even higher than OpenAI's GPT-5.6 Sol, while costing only about one-hundredth as much. This breakthrough is expected to provide a more economical path for the deployment of agentic search.
Neon is a fully managed serverless PostgreSQL database that separates compute and storage, supporting automatic scaling and on-demand scaling to zero. Castform primarily provides AI voice agents and model training platforms for enterprises and developers. The two companies collaborated to post-train an open-source model using reinforcement learning, enabling it to efficiently find the required information in specific data sources.
The Neon team noted that a good AI agent needs two capabilities: finding the correct context data through tools, and routing decisions to determine which model to use for search. The current mainstream approach is embedded search, which converts documents into numerical data and then looks for similar content. However, as agents become more widespread, the search process is shifting toward agentic search, where the model autonomously breaks down problems and performs multiple searches until complete information is gathered.
However, agentic search calls a high-performance model for each search, significantly increasing time and cost. For example, GPT-5.6 Sol takes more than 10 seconds per search request and costs about $0.03. To reduce costs, Castform chose to train a relatively small open-source model to learn how to find the required information from specific data. During the training process, the system scores the model's search results, evaluating dimensions such as correctly finding documents, citing appropriate paragraphs, and the accuracy of the final answer.
In Neon's validation, the Castform post-trained model achieved an average evaluation score of 1.447, surpassing GPT-5.6 Sol's 1.369 and GPT-5.4's 1.377, setting a new record in the validation. In terms of cost, the model's single inference cost is only $0.000929, while GPT-5.6 Sol's inference cost is $0.087338, a difference of nearly a hundred times.
This result shows that small models with targeted training can outperform general large models on specific tasks while significantly reducing inference costs. For AI agents that rely on a large number of search requests, this approach helps reduce latency and fees, making agentic search more practical. As related technologies further develop, this training method may be adopted by more enterprises and developers, driving the intelligent upgrade of document search and agent applications.
