What's this about?
A Neon/Castform benchmark argues that small, purpose-built open models beat frontier models like GPT-5.6 Sol on retrieval tasks when you weigh price and efficiency — roughly 100× cheaper for comparable or better retrieval quality.
What exactly is "retrieval"?
Finding the right information from a body of documents to answer a question — the backbone of RAG (retrieval-augmented generation) and any "chat with your docs" system. It's a narrower skill than general reasoning, and the claim is that you don't need a giant expensive model to do it well.
What's the community saying?
HN 233 pts, 57 comments.
- The harness-design takeaway: a good agent should spin up subagents that offload narrow tasks (like retrieval) to cheap targeted models — mixture-of-models, not one big model for everything.
- Independent corroboration: several commenters report smaller models beating their larger siblings on fact retrieval from documents.
- The harder unanswered question: how does retrieval hold up finding needles in ever-larger haystacks — price is easy, scaling quality is the real test.