Nace AI has introduced Drex for Search, a model designed to optimize decision-making processes for AI agents. The system evaluates potential options using calibrated probabilities in a single forward pass, aiming to improve efficiency in retrieval-augmented generation tasks.
Testing on the OfficeQA Pro benchmark, which includes 133 questions based on U.S. Treasury Bulletins, showed a 6% increase in correct answers. The model also demonstrated a 62% reduction in steps, 56% fewer tokens, and a 2.3x increase in speed compared to previous configurations.
On the BrowseComp-Plus benchmark, which features 830 multi-hop questions across 100 000 web documents, Drex for Search achieved a 92.05% score using 12.6 search calls with GPT-5. Additionally, on the VisDocAgentBench, the system reached a 77.5% R@1 score using an open-weight Qwen 27B planner.
The company reports a 94.7% accuracy rate on the MADQA benchmark using Gemini 3.5 Flash, with a page-attribution F1 score of 88.2. As of October 20, 2026, Nace AI ranks this performance first on the leaderboard.
In the EnterpriseRAG-Bench, which utilizes over 500 000 internal documents, the system achieved an overall score of 88.22. This result places it third on the leaderboard, surpassing OpenAI File Search, which scored 61.03.
