Castform and Neon say a 4B model matched GPT-5.6 Sol at 100x lower cost

Why it matters

Castform's result suggests task-specific post-training can make inexpensive open models competitive on narrow production workflows, though the 100x claim rests on a vendor-authored benchmark with limited methodology.

Castform co-founder Ying Hang Seah and Neon say a 4-billion-parameter open-weight model post-trained with Castform retrieved search results as accurately as OpenAI's GPT-5.6 Sol while costing about 100x less. The claim comes from an August 5, 2026 case study published by Neon and co-authored by Seah, Databricks product executive Pranav Aurora and Angel Pan.

Seah studied computer science and management science and engineering at Stanford, co-founded the real-time code collaboration startup Tym, and later worked as a software engineer at Roblox after its acquisition of Tym.

"Most teams' best training data is just sitting in their databases," Seah said in the Neon case study. Castform's bet is that teams can turn those databases, documents and production traces into specialized models without building an internal machine-learning platform.

A narrow win with useful economics

The benchmark is a vendor-authored, task-specific evaluation. Neon and Castform say their 4B open-weight model matched GPT-5.6 Sol on retrieval accuracy while costing about 100x less per inference workload. The comparison appears in Neon's case study.

Castform and Neon have not disclosed the open-weight model's name or exact checkpoint. They also have not published the complete evaluation corpus, number of test queries, detailed scoring methodology, variance across runs or confidence intervals. Those omissions prevent an independent reproduction of the headline result.

Neon's accompanying text gives one reference point for the cost argument: a typical multi-turn search request using GPT-5.6 Sol takes more than 10 seconds and costs about $0.03 end to end, according to the company. The case study does not specify whether its 100x comparison includes data preparation, synthetic-example generation, reinforcement-learning training, search infrastructure or engineering costs.

The result offers a concrete example of task-specific post-training narrowing the performance gap between a small open-weight model and a much larger API model. Its scope remains limited to the retrieval workflow Neon and Castform evaluated.

Teaching a small model how to search

Castform's system trains a model to decide what to search for, when to search again and how to use the returned evidence. That differs from a conventional retrieval-augmented generation pipeline in which developers perform one embedding lookup and place the resulting chunks into a prompt.

Lakebase Search combines BM25 keyword search through lakebase_text with vector retrieval through lakebase_vector.

During each reinforcement-learning rollout, the model issued search queries and received a reward based on whether it retrieved the correct source, cited the right passage and produced the right answer. Feedback from those attempts adjusted the model's behavior. Castform then used the same search interface during production inference, reducing the difference between the training environment and the deployed system.

Castform and Neon have published a retrieval-training example on GitHub. Castform's documentation describes training from production agent traces, allowing failures and edge cases observed after deployment to become new examples. The example illustrates the workflow but does not provide enough information to reproduce the Neon benchmark.

Neon's role becomes more important during training, when thousands of parallel rollouts may each make repeated database calls. Neon says its separation of storage and compute lets the system scale up for those bursts and scale down between them. Castform can reuse Postgres as the document store, keyword index, vector index and training-time search service instead of maintaining a separate system for each job.

Seah's bet on specialized models

Castform entered open beta on June 11, 2026, less than two months before publishing the Neon result.

Castform asks developers to define the task, the tools available to the model and the reward signal that represents a good result. Castform handles dataset generation, GPU provisioning, reinforcement-learning algorithms, checkpointing and monitoring. Its documentation also supports training from production agent traces, allowing failures and edge cases observed after deployment to become new examples.

The product is self-service and usage-based. Castform's pricing page lists Qwen3.5 4B training at a launch rate of $0.10 per 1 million training tokens. This is a listed training rate, rather than the total cost of producing the model in the Neon benchmark, and Castform has not identified Qwen3.5 4B as the model used there. Castform also has not disclosed the benchmark's number of training tokens, rollouts or data-generation calls.

Castform is entering a young post-training market that includes OpenPipe, Predibase and Prime Intellect. OpenPipe, Predibase and Prime Intellect share a premise: many production agents need reliable performance on one workflow with a known set of tools and proprietary data, rather than broad improvements across coding, reasoning and knowledge.

Seah's test gives that premise a concrete retrieval example. The reported match with GPT-5.6 Sol supplies a price and performance reference, although the undisclosed methodology and narrow task do not support claims about replacing Sol across other retrieval workloads.

For Castform, that boundary fits the product strategy. Seah is selling control over a specific behavior, including the ability to export the trained weights and deploy them elsewhere. Dependable training loops could let teams reserve frontier models for work requiring broader capabilities while moving repetitive, high-volume agent steps onto models they control.