Google has rolled out a major upgrade to Gemini Deep Research, adding two new agents: Deep Research and Deep Research Max. Both now run on Gemini 3.1 Pro. The standard version is tuned for lower latency and lower cost, while Max is designed for deeper research output. The products have entered public preview through paid Gemini API plans, with a Google Cloud release set to follow.
Deep Research Max spends more compute time to produce deeper reports
The main difference in Deep Research Max is what Google calls “extended test-time compute.” Instead of running one pass and returning an answer, the agent repeatedly reasons, searches, and revises before delivering a report. Google said Max shows a sharp improvement in retrieval and reasoning compared with the preview released in December, and that it now draws from a much larger set of sources during research.
When evidence conflicts, Max can cite authoritative material such as SEC filings and peer-reviewed journals. Google framed the product as a tool for longer-running assignments: a team can schedule a task overnight and arrive in the morning to a completed due diligence report. Speed is not the goal here. Depth is.
MCP support expands the agent from web search to private data systems
Another major change is native support for MCP, or Model Context Protocol. Earlier versions of this type of agent were mostly limited to public web information. With MCP, Deep Research can connect to custom enterprise sources and professional data feeds, letting users combine public internet content, private APIs, and internal databases inside one workflow.
Google said FactSet, S&P Global, and PitchBook have worked with the company on MCP servers so customers can pull financial and market data from those platforms directly into Deep Research. For finance teams and research firms, that creates a single path across internal ERP systems, market-data providers, and company databases without moving manually between separate tools.
Users can enable Google Search, remote MCP, URL Context, Code Execution, and File Search together. They can also switch off network access entirely and keep the agent operating only inside approved databases, a setting aimed at organizations with strict concerns around data leakage.
Native charts, planning controls, and live streaming are built into the workflow
Google also introduced three product features tied directly to research work. The first is native charts and infographics. According to the company, this is the first time Gemini API supports this kind of output, allowing Deep Research to generate HTML charts or Nano Banana infographic assets instead of returning only text.
The second is collaborative planning. Before starting a research job, the agent produces a research plan that users can inspect, adjust, and guide. That gives teams more control over scope and structure before execution begins, rather than treating the system as a black box that accepts a prompt and returns a report.
The third is live streaming. The system surfaces summaries of intermediate reasoning steps so users can see what the agent is doing while a long task is still running. Text and images are streamed as they are generated, giving more visibility during extended research sessions.
Deep Research also now accepts multimodal inputs including PDFs, CSV files, images, audio, and video. That reduces the need for manual preprocessing when analysts are working across different file types.
Google is pushing AI agents closer to due diligence work
Based on the feature set Google described, Deep Research Max is moving beyond basic search and summarization. Repeated reasoning, handling conflicting evidence, citing regulatory filings, and accessing private financial datasets through MCP place it much closer to the kind of due diligence support normally associated with junior analysts.
Google did not present that as full replacement. Questions around validating the agent’s reasoning, controlling permissions for private data, and using AI-generated research conclusions in regulated settings still remain. What the company made clear is narrower but important: the technical path for enterprise testing is now open.