Google announced that the newly upgraded Gemini Deep Research agent will be available to developers through the new Interactions API. The agent uses Gemini 3 Pro as the core reasoning model, specifically for long-term, multi-step research tasks, and generates structured and sourced research reports for users through automatic planning and querying, repeated retrieval and sorting of information.
According to the official introduction, Gemini Deep Research has achieved the latest results on a number of deep search benchmarks: 46.4% on the full set of Humanity's Last Exam, 66.1% on DeepSearchQA, and 59.2% on BrowseComp, all of which are better than the performance of the basic Gemini 3 Pro model. Developers can use the Interactions API to call the agent to perform background long tasks in the application, and combine web search with their own data sources to achieve automated data collection, comparison, and summary.
Currently, the Gemini Deep Research agent is still in preview and can only be called by specifying the agent name through the Interactions API, which supports working with file retrieval, streaming output, and other capabilities. Google also announced that it will gradually introduce this proxy to products such as Gemini Apps, Google Search, NotebookLM and Google Finance to provide end-users with stronger in-depth research capabilities.
FAQ
Q: What exactly is Gemini Deep Research agent?
A: Gemini Deep Research is a "deep research" agent launched by Google that can automatically plan multi-step searches, read web pages and user documents, and output structured research reports.
Q: What is the difference between Gemini Deep Research and the regular Gemini 3 Pro model?
A: They share Gemini 3 Pro as the core model, but Deep Research has added a dedicated agent workflow to train and enhance long-term retrieval, reduce hallucinations, and improve report quality.
Q: How can developers call Gemini Deep Research in the API?
A: Developers need to use the Gemini API key to create an interaction session through the Interactions API, specify the Deep Research agent name, and set the background parameter to true to perform the background long task.
Q: How did Gemini Deep Research perform in benchmarks?
A: The official disclosure has achieved the latest leading results in three in-depth research benchmarks, Humanity's Last Exam, DeepSearchQA and BrowseComp, which is significantly better than the foundation model without an agent process.
Q: When will the average user be able to experience Deep Research in a consumer product?
A: Google said it will integrate Gemini Deep Research into Gemini App, Search, NotebookLM and Google Finance products "soon", and the specific region and account opening rhythm are subject to Google's subsequent announcements.
