Castform trained a 4-billion-parameter open-weight model that matches GPT-5.6 Sol on retrieval accuracy at one hundredth the cost. The model, post-trained with reinforcement learning on Neon infrastructure, retrieves search results with the same precision as OpenAI’s frontier.

Why this matters to you. If you’re building RAG pipelines or agentic systems with search, the economics shift. A multi-turn search request with GPT-5.6 Sol takes over 10 seconds and costs roughly $0.03 (source: Neon, August 5, 2026). A specialized small model costs 100 times less. As we reported on August 1st, DeepSeek V4-Flash had already closed the gap with GPT-5.6 Luna on cost per task. Now the pattern repeats on a specific domain: retrieval.

On specific tasks like searching well through a corpus, a 4B model post-trained with RL on your data is more than enough. The frontier remains necessary where you need generality, not where you need retrieval. Castform transforms your existing corpus into training data: documents, wikis, support tickets become synthetic questions and reward functions. If you have a company knowledge base, your training dataset already exists.

If you want to try it. The Neon post describes the end-to-end pipeline: start with your corpus on Postgres, let Castform generate questions and rewards from the text, and train a model that in production uses the same search tool it learned during training.

In detail

The piece describes a two-phase shift you’ve seen elsewhere, but compressed into a single announcement here.

What came before. In 2022 the industry went all-in on embeddings: every database added vector search, and pgvector became Neon’s most-downloaded extension. Classical RAG pipelines ran a single similarity query and passed the result to the model. By 2025 agents started doing multi-hop search: decompose a big problem into sub-queries, plan, search multiple times in a loop. Each loop iteration was a call to the frontier model. Result: latency and cost per request went up. The Neon post quantifies the price of this approach with GPT-5.6 Sol at over 10 seconds and roughly $0.03 per request (source: Neon blog, August 5, 2026).

What changes. Medium-small open-weight models cost an order of magnitude less, but out of the box they’re less capable on agentic tasks. Post-training with reinforcement learning closes the gap on specific tasks. Castform sits here: making post-training RL as accessible as prompt engineering, without managing GPUs and machine learning internals.

How the pipeline works. Castform starts with the company corpus living on Postgres (Neon with Lakebase Search). It uses the lakebase_text and lakebase_vector extensions to generate synthetic data: given a document, it infers ground truth, creates a question that document should answer. The document becomes the task, answer correctness becomes the reward function. The model attempts the search, the score guides it to improve. In production it uses the same search tool it learned during training.

The clever move is that Castform solves the training data problem, which is the real bottleneck. Companies have proprietary corpora (internal docs, product records, support articles, customer interactions) but no clean datasets with tasks and rewards ready to go. Castform automates that transformation.

Limitations. This is a promotional post from Neon and Castform, not a peer-reviewed paper. The accuracy benchmark is specific to their retrieval setup: we don’t know how it generalizes to other domains or corpus types. The model specializes in search, it doesn’t replace a general model. The 100x savings are calculated on their specific use case (multi-turn search with GPT-5.6 Sol) and depend on current API pricing, which changes. Castform’s actual availability, training costs, and hardware requirements aren’t detailed in the post.

For anyone wanting to measure on their own whether a small post-trained model holds up against their reference frontier on their own case, the playbook Compare two models in fifteen minutes gives you the structure: same task, same test cases, results table.