Editorial illustration for Nace AI's Drex posts 92.05% on BrowseComp-Plus with 12.6 search calls
Drex Hits 92% Accuracy on Complex Search Tasks
Nace AI's Drex posts 92.05% on BrowseComp-Plus with 12.6 search calls
Nace AI put a number on a problem most retrieval-augmented generation setups quietly dodge: finding the right answer across a million-plus files when that answer sits in a scanned PDF, a table buried on page 23, or requires hopping between two documents. The company's new release, Drex for Search, scored 92.05% on the BrowseComp-Plus benchmark when paired with GPT-5, averaging 12.6 search calls per question. It's built as a hosted retrieval layer, deployable on Nace's managed cloud, a dedicated instance, a customer's own AWS or Azure account, or fully air-gapped on-prem.
The system pairs with answering models like GPT-5, Gemini 3.5 Flash, or Qwen 27B, and coding agents including Claude Code, Codex, Cursor, and OpenCode can call it directly through CLI, SDKs, MCP, or REST. Nace says it now ranks first or second on three of four public leaderboards, with a standout 77.5% R@1 score on VisDocAgentBench, ten points ahead of Claude Opus 5's 67.5%.
The pitch rests on a specific critique of standard RAG pipelines, chunk, embed, top-k, hope, and how Nace's Drex decision models are meant to replace that guesswork.
Nace AI has released Drex for Search, which brings its Drex decision models to document search. Drex for Search is a retrieval layer that helps people and agents find evidence. Coding agents such as Claude Code, Codex, Cursor and OpenCode call it through CLI, SDKs, MCP or REST.
Why this matters
The call-efficiency number is the real story here, not the accuracy score. Sitting at 92.05% while using 16.6x fewer search calls than the top entry suggests Drex for Search is optimizing for something developers actually pay for: latency and API cost at scale. Any agent making 209 search calls per question might win a leaderboard, but it's not going into production for a team watching their OpenAI bill. That tradeoff is what founders building retrieval-augmented agents should care about.
We'd want to see how Drex performs outside BrowseComp-Plus's fixed 100K-document corpus before treating 92.05% as gospel. Multi-hop benchmarks over static web documents don't always predict behavior against a live, 1M+ file index with messier formats and duplicate content. The CLI, SDK, MCP and REST hooks into Claude Code, Codex, Cursor and OpenCode matter too.
If Drex plugs into existing agent workflows without a separate integration project, that's the adoption lever, more than the leaderboard rank. Watch for independent benchmarks on real production document sets, not just the curated eval set Nace chose to report.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Tsinghua AI Open-Sources AKV Cache Server for LLM Agents on SGLang
From Archives Nvidia open-sources Flora, robot video labeling 125x cheaper than GPT-6 Astra
Common Questions Answered
What is Drex for Search and how does it improve retrieval-augmented generation?
Drex for Search is a hosted retrieval layer developed by Nace AI that helps find evidence across large document collections by applying Drex decision models to document search tasks. It addresses the challenge of locating answers buried in scanned PDFs, tables, or across multiple documents by optimizing both accuracy and search efficiency for production deployments.
What benchmark score did Drex for Search achieve on BrowseComp-Plus?
Drex for Search scored 92.05% on the BrowseComp-Plus benchmark when paired with GPT-5, while averaging only 12.6 search calls per question. This performance demonstrates a significant efficiency advantage, using 16.6x fewer search calls than competing top entries on the leaderboard.
Why is call-efficiency more important than accuracy for Drex for Search in production environments?
Call-efficiency directly impacts latency and API costs at scale, which are the primary concerns for development teams deploying retrieval-augmented agents. While competitors may achieve higher accuracy scores with 209+ search calls per question, this approach becomes economically impractical for production use due to cumulative OpenAI API expenses.
What deployment options does Drex for Search provide for developers?
Drex for Search is deployable as a hosted retrieval layer on Nace's managed cloud infrastructure and can be accessed by coding agents through multiple interfaces including CLI, SDKs, MCP, and REST APIs. This flexibility allows integration with various AI agents such as Claude Code, Codex, Cursor, and OpenCode.
What types of documents can Drex for Search help locate answers within?
Drex for Search is designed to find answers across million-plus file collections, including scanned PDFs, tables buried within documents, and information that requires cross-referencing between multiple documents. This capability addresses a significant gap in retrieval-augmented generation systems that typically struggle with complex document navigation scenarios.
Further Reading
- Drex for Search: Supercharge Your Agentic Search - Nace AI
- BrowseComp-Plus: A More Fair and Transparent Evaluation of Deep-Research Agents - arXiv
- BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agents - GitHub
- Nace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls - MarkTechPost
