The agent takes an iterative approach to research: it formulates search queries, reads results, spots knowledge gaps, and searches again until it reaches a satisfactory answer. The new version is more powerful than the previous model and slightly outperforms the web search capabilities of the recently released Gemini 3 Pro.
Google says the agent's reasoning core runs on Gemini 3 Pro, trained specifically to reduce hallucinations and maximize report quality for complex tasks. AI errors remain a weakness across all current deep research systems—while you can't fully trust the results, Deep Research can still prove useful for exploratory source gathering.
New benchmark scores show where Google stands
Google claims the new deep research agent hits state-of-the-art results on several benchmarks: 46.4 percent on the full Humanity's Last Exam (HLE), 66.1 percent on the new DeepSearchQA, and 59.2 percent on BrowseComp. The agent is also optimized to produce well-researched reports at lower cost.
Google's Deep Research Agent leads on Humanity's Last Exam and DeepSearchQA, but trails GPT-5 Pro slightly on BrowseComp. | Image: Google
