Google has released a new version of Gemini Deep Research. This is an agent designed to automate complex research tasks.
The agent runs on Gemini 3 Pro. The model can process handwriting, graphs, and mathematical notation. It incorporates this visual information directly into reports and search queries. As a result, the system can not only search textual sources, but also retrieve data that was previously difficult to automate, according to SiliconANGLE.
The agent works iteratively and plans its research steps independently. It formulates search queries, evaluates the results found, and determines where information is missing. Google states that the web navigation in this release is better able to penetrate deep into sites. Users can upload documents, which are then automatically scanned for relevant passages. The agent can summarize, interpret, or combine these documents with information from public sources.
Interactions API as a central access point
Deep Research is now available through the new Interactions API. This API acts as a central access point for both Gemini models and pre-built agents. Google plans to add additional agents and offer support for custom agents in the future. The API also takes some of the data management off your hands, which means developers spend less time processing and structuring files. Models can also be linked to external systems via MCP.
Google reports that the agent performs better on benchmarks than previous versions. On Humanity’s Last Exam, a benchmark with more than 2,500 questions focusing primarily on mathematics, physics, and programming, the agent scored 46.4 percent.
On DeepSearchQA, a dataset released by Google, the system also outperformed its predecessors. This benchmark consists of multi-step information tasks that depend on previous analyses and is designed to measure both precision and completeness. Google also uses the dataset internally to investigate how additional reasoning steps and longer processing time affect performance.
According to reports from SiliconANGLE, Google Deep Research is primarily positioned as a tool for sectors that involve a lot of document analysis and source collection. In that context, the agent is mainly intended to take over repetitive and time-consuming tasks, while the Interactions API simplifies integration into applications. With the introduction of this version, Google wants to show how language models are increasingly capable of handling research tasks that go beyond traditional search or summarization functions.