Google has rolled out two autonomous research agents, Deep Research and Deep Research Max, built on Gemini 3.1 Pro. The product shift is clear: instead of using AI as a chat interface that answers one prompt at a time, users can now hand off an entire research workflow. The agent first discusses a research plan, waits for approval, then runs in the background for as long as 60 minutes before delivering a report that can include charts.
How Deep Research and Max differ
The standard version is designed for speed. According to Google, Deep Research runs about 80 searches per task, making it better suited for faster turnaround and more interactive use. Deep Research Max is aimed at broader and deeper analysis, with roughly 160 searches per task. It is positioned for longer-running assignments that users may leave running overnight.
Google said the Max version repeatedly refines its report and described the approach as “extended test-time compute”. It also consumes much more context than the standard version, using about 3 to 4 times as many tokens, or around 900,000 input tokens and 80,000 output tokens. Pricing follows Gemini 3.1 Pro token rates: $2.00 per million input tokens and $12.00 per million output tokens. On that basis, Google estimates a standard task at about $1 to $3, while a Max task comes in at roughly $3 to $7.
MCP support brings in enterprise databases
One of the most important additions is support for Model Context Protocol (MCP), which allows the agents to connect to internal company data sources. Google said it is working with FactSet, S&P Global, and PitchBook on MCP integrations. For financial due diligence work, that means the agent can pull from company filing databases and other professional data sources rather than relying only on the public web.
Google also said web search can be turned off entirely, allowing a task to run only against internal data. That changes the tool from a web research assistant into something closer to a controlled enterprise research system.
Charts are generated inside the report
The agents can produce native HTML charts and infographics directly in their reports. This removes an extra step that often sits between gathering data and presenting it, where users have to export numbers into spreadsheet tools to build visuals manually. The result is still editable, which makes it usable as a working analysis file rather than a fixed text output.
Google also built the workflow around collaboration. The agent does not start immediately after receiving a question. It first drafts a research plan for review, and users can revise or approve it before execution. If the final report misses the mark, they can continue with follow-up prompts and redirect the task through multi-turn conversation.
Public preview access is live
Both Deep Research and Deep Research Max are now in Public Preview and available through the paid Gemini API tier. Google AI Studio already supports direct use, and both the Python SDK and JavaScript SDK are available.