A 4B open model matches GPT-5.6 Sol on retrieval at 1/100 the cost
Neon and Castform's runnable recipe swaps a frontier retrieval API for weights you own — plus an unpatched Rovo bug that leaks Jira and Confluence data.
Copy markdown
Frontier-grade retrieval from a model 100x smaller
Neon and Castform post-trained a 4B open model with reinforcement learning until it matched GPT-5.6 Sol's retrieval accuracy on their own benchmark — at roughly one-hundredth the cost. For anyone running RAG, that's the retrieval leg of the pipeline getting an order of magnitude cheaper without trading away answer quality.
The cost math: ~$0.03 and 10s a query, gone
By their numbers, a multi-turn search with GPT-5.6 Sol runs over 10 seconds and about $0.03 end-to-end per request; the tuned open model does the same work faster and far cheaper. At scale that's the gap between retrieval being a rounding error and a real line item on your bill.
Try it this weekend: the neon_rag example is live
The recipe ships as a runnable GitHub example (castform-ai/benchmax, examples/neon_rag): Neon's Lakebase hybrid search — BM25 plus vector, fused with reciprocal rank fusion — as the retrieval tool, and Castform's RL platform doing the post-training. You bring a corpus and walk away with weights you own instead of a metered API call.
Elsewhere: Atlassian Rovo leaks your data, unpatched
PromptArmor showed Rovo can be steered by a single poisoned document into building URLs that ship your Jira tickets and Confluence pages to an attacker's server — and it still fires with web search turned off. Flagged to Atlassian back in May, still unpatched as of Aug 5; until it's fixed, keep untrusted content away from Rovo.