Glonce

Castform claims RL post-trained open models beat GPT-5.6 Sol on retrieval

agentsThe Glonce editorsHacker News🔥74Glonce rating74Impact60Novelty60Relevance85Surprise55Players45Credibility62
Castform claims RL post-trained open models beat GPT-5.6 Sol on retrieval

Neon's blog published a post, written from Castform's point of view, arguing that small open-weight language models, once trained with reinforcement learning (RL) post-training, can match or beat GPT-5.6 Sol on retrieval tasks while running far cheaper. The post frames a good search agent as needing two things: context, meaning tools to find the right data, which it says Neon's Postgres service (which the post calls Lakebase) and its new Search extensions supply, and a model that knows what to search for, which is what Castform, a startup offering RL post-training as a service, is built to provide.

The post traces what it calls the evolution of agentic search. Around 2022 the industry converged on embedding-based similarity search, and pgvector became Neon's most downloaded extension as engineers built RAG pipelines by hand. Around 2025, agentic search took over: instead of issuing one query, models plan and search repeatedly in a loop, decomposing a problem into smaller pieces across multiple hops, and every loop iteration is another call to a frontier model. The post gives a concrete number for what that costs: a typical multi-turn search request run with GPT-5.6 Sol takes more than 10 seconds and costs about $0.03 end to end, which it calls prohibitively slow and expensive. Small open-weight models are already about 100 times cheaper than closed API models, the post says, but out of the box their capabilities lag; RL post-training is pitched as the way to close that gap, so that on tasks like search, post-trained open-source models can match or beat frontier models while costing orders of magnitude less per request, though no specific dollar or latency figure is given for a post-trained model itself.

Castform's pitch is to make RL post-training as approachable as prompt engineering, without a developer having to handle machine learning or GPU infrastructure directly. Effective post-training needs three things: a task, such as answering a user's question; an environment for the agent, such as a search tool over the company's own data; and a reward function that scores whether an attempt succeeded, such as whether the answer is correct. The model then repeats the task, gets scored, and adjusts toward better performance. The post argues most companies already hold the raw material for this inside their own systems, internal documentation, product records, support articles, customer interactions, wikis, and operational databases, but lack a clean set of ready-made tasks and reward functions, which is why, per the post, many teams dismiss post-training, saying they do not have the training data, or that fine-tuning is too difficult and requires infrastructure they do not have. Castform's answer is to turn an existing corpus into training tasks itself. As a worked example, the post builds a training case from a travel-policy document: train bookings made through Navan are paid with a GitLab travel card, in standard cabin class, with a 14-day booking lead time required; from that document Castform infers a ground truth answer and synthesizes a natural-language question, without stating that GitLab or Navan is an actual customer of either company. Castform also gives visibility into the training run itself, letting a developer watch the reward climb step by step and drop into individual tasks to debug problems such as broken tools or reward hacking.

The post's second half explains why it says Neon fits this job. During training, the agent calls Lakebase Search repeatedly until it has enough context to answer, and across thousands of parallel rollouts, each potentially making dozens of calls, that adds up to a highly bursty workload. Neon's dynamic compute scaling is said to absorb those peaks without Castform having to provision for maximum capacity around the clock, so a training run gets low-latency search when demand spikes while compute scales down during idle periods. The post adds that this matters more once agents move beyond search into modifying data: training a stateful agent needs isolated environments that can be created and reset cheaply, so one rollout's actions cannot affect another or touch production. Neon branching is said to give each rollout its own isolated database state, and time-travel queries let that state be reconstructed and inspected afterward, which the post frames as a path toward training thousands of stateful agent rollouts without keeping thousands of environments running continuously.

The post quotes Castform cofounder Ying Hang Seah on the rationale for pairing the two products: most teams' best training data is already sitting in their own databases, turning that raw data into something usable is hard, and letting agents read, search and mutate data cheaply at scale requires infrastructure most teams do not have; pointing Castform at Neon is presented as a way to skip both problems at once. The post closes with a direct call to post-train a first model at castform.com. On Hacker News, where it was submitted under the title Beating GPT-5.6 Sol on retrieval with 100x cheaper open models, the discussion had drawn 251 points and 64 comments about 10 hours after posting.

Key facts

  • Castform, a startup offering reinforcement learning (RL) post-training as a service, argues in a post on Neon's blog that once RL post-trained, small open-weight models can match or beat GPT-5.6 Sol on retrieval tasks.
  • The post says a typical multi-turn search request run with GPT-5.6 Sol takes more than 10 seconds and costs about $0.03 end to end; small open-weight models are already about 100 times cheaper than closed API models before any post-training, though no specific dollar figure is given for a post-trained model.
  • Castform turns a company's existing documents, wikis and support articles into RL training tasks with a synthesized question, an inferred ground truth answer and a reward function, then gives observability into the training run to debug problems such as broken tools or reward hacking.
  • Castform runs its RL training loop against Neon's Lakebase Search, leaning on Neon's autoscaling to absorb bursty traffic from thousands of parallel rollouts, each making dozens of search calls, and on Neon branching plus time-travel queries to isolate and inspect each rollout's database state.
  • On Hacker News, the post had drawn 251 points and 64 comments about 10 hours after posting.

Why it matters

Agentic search means a model searches in a loop rather than issuing one query, and each loop iteration is another call to a frontier model, which the post says makes a typical multi-turn search request with GPT-5.6 Sol take more than 10 seconds and cost about $0.03 end to end. RL post-training is pitched as a way out of that trade-off: instead of paying frontier prices for every search step, a much cheaper open-weight model is trained specifically on the retrieval task it needs to do well, closing the capability gap that otherwise keeps cheap models out of production agent loops. If the pitch holds up in practice, it changes the calculus for anyone building search or retrieval agents around per-call frontier API costs, and it doubles as a case for Neon's own database features, positioned here as the infrastructure such training runs need.

Who it affects

Developers and companies building agentic search or retrieval-augmented systems who are currently paying frontier-model prices for repeated search calls, especially those sitting on a large body of proprietary data, internal documentation, product records, support articles, customer interactions, wikis, and operational databases, that could become RL training material. It also concerns existing Neon customers already running Postgres-based search or RAG pipelines, since the pitch is built specifically around Neon's Lakebase Search, branching and autoscaling features rather than being infrastructure-agnostic.

How to use it

Per the post, a developer would point Castform at their own Neon database through Lakebase Search, then define a task such as answering a user's question, an environment such as a search tool over their corpus, and a reward function that scores whether the retrieved chunks, cited sources and final answer are correct. Castform turns existing documents into synthetic question-and-answer training data and runs the RL loop, with observability into reward progress and individual task traces to debug problems such as broken tools or reward hacking. The post links out to a code example and an example training run, and closes with a direct call to post-train a first model at castform.com; no subscription price or usage tier for Castform's service is given in the article.

How solid is it

The central claim, that RL post-trained open-weight models can match or beat GPT-5.6 Sol on retrieval, is stated in general terms and is not backed in the text by a named benchmark, dataset or accuracy score; no side-by-side test of a Castform-trained model against GPT-5.6 Sol on a specific task is described. Hard numbers appear only for the frontier baseline, more than 10 seconds and about $0.03 per multi-turn search request with GPT-5.6 Sol, and for the pre-training cost gap, open-weight models about 100 times cheaper than closed API models; no dollar figure, latency number or retrieval-quality score is given for a Castform-trained model, whose advantage is described only as orders of magnitude less per request. The piece is also self-published: it runs on Neon's own blog, is written from Castform's point of view, and both companies have a commercial interest in the pairing looking credible, closing with a direct link to castform.com. Hacker News reception, 251 points and 64 comments about 10 hours after posting, shows engagement with the claim, not independent verification of it.

Risks and caveats

The article does not name which open-weight model family Castform post-trains, so it is not possible to tell from the text what is actually being compared to GPT-5.6 Sol. No pricing or subscription terms for Castform's service are given. The travel-policy passage built from a Navan booking paid by a GitLab card is presented only as a worked example of turning a document into a training task; the article does not say either company is an actual Castform or Neon customer. Neither the blog post's publication date nor GPT-5.6 Sol's release date is given, so how current the cost and latency comparison is stays unclear. The thousands of parallel rollouts and dozens of calls per rollout describing training load are stated in the post as approximations, not measured counts.

“Most teams' best training data is just sitting in their databases. The problem is that turning raw data into something usable is hard, and letting agents read, search, and mutate data cheaply at scale requires advanced infra. Pointing Castform at Neon skips both.”

— Ying Hang Seah, cofounder of Castform