Nace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls

Nace AI has released Drex for Search, a hosted retrieval layer that coding agents like Claude Code, Codex, Cursor and OpenCode can call through a CLI, SDKs, MCP or REST. Nace reports 92.05% accuracy on the BrowseComp-Plus benchmark with GPT-5, using about 12.6 search calls per question, which it ranks second of 86 entries.

The pitch is efficiency. The only entry ahead of it on that leaderboard used 209.27 calls per question, so Nace is claiming nearly the same accuracy with a fraction of the work. For teams paying per token and per agent step, that is the number that matters.

We like the problem Nace is going after. The classic retrieval setup it criticizes, chunk the documents, embed them, grab the top matches and hope, really does fall over when the answer sits in a table deep in a PDF, needs two hops across files, or lives in a scan. Anyone who has watched an agent burn dozens of turns hunting through a folder knows the pain. A tool that narrows from folder to file to page to table, then hands only that to the model with a file-and-page citation, is the kind of plumbing that makes agents more useful and easier to check.

The citations are the part we'd highlight most. An answer that names the exact page, row or minute it came from can be verified by a human. That is how agents earn trust in work like tax, contracts and SEC filings, which are among the domains Nace lists.

Now the caveats, and they are significant. The write-up we read is labeled as sponsored by Nace AI, and the rankings and scores are as published by Nace, from leaderboards it read on October 20. That does not make the numbers wrong, but nothing here is independent replication. Nace's own table also shows that Drex did not help every pairing: on OfficeQA Pro, Codex with GPT-6 luna dropped from 54.9% to 53.4% accuracy, even as agent turns fell for all nine pairs tested.

There is also a dependency question. Drex is a hosted system, though Nace says it can also run in a customer's own AWS or Azure account, a dedicated cloud, or on-prem and air-gapped, with zero retention available when customers bring their own storage. Teams putting sensitive documents through any retrieval layer should treat those deployment options as the real decision, not the leaderboard rank.

Our read: promising, focused on a real bottleneck, and unproven outside the vendor's own reporting. I'd want third-party runs before betting a workflow on it.

BrowseComp-Plus asks a deep-research agent 830 multi-hop questions over about 100,000 fixed web documents. Nace also reports 77.5% recall-at-one on VisDocAgentBench with an open-weight Qwen 27B planner versus 67.5% for Claude Opus 5, 94.7% accuracy on MADQA with Gemini 3.5 Flash, and an 88.22 score on EnterpriseRAG-Bench, ranked third, against 61.03 for OpenAI File Search.