Google just launched two new autonomous research agents, Deep Research and Deep Research Max, both built on its Gemini 3.1 Pro model. They're available now in public preview for the Gemini API's paid tier via the Interactions API, with plans to roll out to Google Cloud for startups and enterprise customers later.
Deep Research is optimized for lower latency and cost, replacing the initial preview version from last December. It's designed for scenarios where you'd plug it directly into a conversational interface. Deep Research Max is the new, more powerful configuration. It throws more computational muscle at a problem, repeatedly searching and refining its reports. Google's suggested use case? Set it up to run overnight on a cron job, and your analyst team wakes up to a due diligence report on their desks.
This new version packs four major upgrades. First, it integrates with the Model Context Protocol (MCP), letting you hook up third-party data sources like financial or market data feeds, or custom tools. Second, it can now natively generate charts and infographics, using HTML or Nano Banana for visualization—no more text-only outputs. Third, it supports PDFs, CSVs, images, audio, and video as multimodal context. Fourth, you can now review and edit the agent's proposed research plan before it executes. You can run Google Search, URL Context, Code Execution, File Search, and MCP tools all at once, or you can completely shut off web access and have it search only within your private data. A real-time streaming interface pushes out intermediate reasoning steps along with text and image outputs.
Google is also working with FactSet, S&P Global, and PitchBook to build out MCP services, allowing those companies' clients to pipe their proprietary data directly into the Deep Research workflow. The official blog post didn't provide specific benchmark scores, but it claims Deep Research Max "consults significantly more sources" than the December version and can "identify key details the old version often missed." This push for more thorough, automated analysis comes as competitors like Anthropic are also enhancing their models for complex tasks, as seen with the recent launch of Claude Opus 4.7.