Beating GPT-5.6 Sol on retrieval with 100x cheaper open models

By Pranav Aurora · Neon (2026-08-05) · On Hacker News (2026-08-06)

359 points · 87 comments on HN · read original →

Points and comments are a snapshot, not live.

Castform and Neon combine post-training and scalable search to beat frontier models at retrieval for 100x less cost.

Castform uses RL post-training to turn proprietary data into task-specific open-weights models that match or beat frontier models like gpt-5.6-sol on retrieval, at 100x lower cost. Neon provides Lakebase Search (hybrid BM25 + vector) and dynamic compute scaling to handle bursty training and inference workloads. The system generates synthetic question-answer pairs from existing corpora, then runs a trial-and-error RL loop with a reward function that scores retrieval accuracy, citation correctness, and answer correctness. Neon branches can isolate each training rollout, enabling stateful agent training without provisioning thousands of environments.

A concrete example: gpt-5.6-sol takes >10s and ~$0.03 per multi-turn search request, while the Castform approach claims comparable accuracy at orders-of-magnitude lower cost. The authors argue that most teams already have the data (internal docs, support articles, wikis) but lack the infrastructure to turn it into training tasks.

What commenters are saying

Commenters broadly see opportunity in specialized, purpose-built models that offload specific tasks from frontier models, reducing cost and latency. Several point to existing examples (Chroma's Context1, SID-1, ZeroEntropy distillation) and note that routing to smaller models is already happening (Claude Code's Explore agent, now capped at Opus).

A counter-current argues that specialized models rarely beat strong general models, and that for repetitive tasks a programmatic solution may be better than delegating to another model. Castform founders engage directly, arguing that proprietary data gives specialized models an edge and that fine-tuning pipelines are trivial to rerun on new base models. Some note that frontier labs are incentivized against this direction because it pulls intelligence to local compute.