A 4B open model matches GPT-5.6 on search at 100x less cost

Neon describes how it took a 4-billion-parameter open-weight model and post-trained it with Castform, a reinforcement learning platform, until it matched the frontier model GPT-5.6 Sol on an agentic search task. The task is the kind that shows up constantly in retrieval systems: break a question into several search steps across a document corpus, then pull the specific clauses that answer it. Their example is a company policy lookup, where the agent has to find the right passage in internal documentation rather than guess.

The pitch of Castform is that it makes post-training feel closer to prompt engineering than to a research project. It automates the RL loop of running a task, scoring the result, and feeding that signal back. On cost, Neon puts hard numbers on the frontier baseline: a typical multi-turn search with GPT-5.6 Sol takes more than 10 seconds and costs about $0.03 end to end. The tuned 4B model, they say, does comparable work for roughly a hundredth of the price.

The claim comes with caveats worth reading closely. This is a vendor post promoting Neon's own platform, it does not give an exact accuracy percentage, and it does not name the base open model. The numbers are specific, but the comparison is theirs to define.

Why it matters

If you run retrieval-heavy agents at scale, a small model tuned for one narrow task can cut inference cost sharply, but this is a vendor benchmark. Test it on your own corpus and your own accuracy bar before you trust the 100x figure.