Google Launches Deep Research Agent: Quickly Get Started with Automated 160 Search Runs to Generate Charts, Claiming to be an AI for Research Collaboration

Google AI officially launches Deep Research and Deep Research Max, two autonomous research agents built on Gemini 3.1 Pro. They can automatically plan research paths, perform up to 160 web searches, access enterprise internal databases (via MCP), and directly generate charts within reports. Moving from “chatting with AI” to “letting AI run complete research for you,” this is an upgrade in Google’s definition of AI agents.
(Background recap: Google upgrades Gemini Deep Research Max: integrates MCP to connect to internal enterprise databases, native charts, enabling analysts to conduct due diligence.)
(Additional background: Jensen Huang sends company-wide email embracing OpenAI Codex: over 10k NVIDIA employees are already using it, GPT-5.5 runs on GB200.)

If you’ve used Gemini’s general chat function for research, you probably know its limits—asking one question yields one answer, searches are limited to a few runs, and you have to piece together conclusions yourself. The Deep Research series does things differently; it will first discuss the research plan with you, and once you agree, it will run autonomously. The background execution can last up to 60 minutes, and it will deliver a complete report with charts.

Google is releasing two versions this time: Deep Research and Deep Research Max, differing in speed and depth.

How to choose between the two versions

Deep Research is the speed-priority version, running about 80 searches per task, suitable for real-time interaction and quick queries. You input a question and get results within a few minutes.

Deep Research Max is the comprehensive version, running about 160 searches per task, with token consumption 3-4 times that of the standard version (approximately 900k input + 80k output tokens). It’s ideal for overnight deep analysis tasks. It will repeatedly refine the report, which Google calls “extended test-time compute.”

In terms of cost, the standard version costs about $1-3 per task, while Max costs about $3-7. Billing is based on Gemini 3.1 Pro’s token rates: $2.00 per million input tokens, $12.00 per million output tokens.

Three key features

MCP access to enterprise databases: Through Model Context Protocol (MCP), you can connect Deep Research to your own data sources. Google is collaborating with FactSet, S&P Global, PitchBook to develop MCP integration, enabling financial due diligence. Users can have the agent query corporate financial report databases directly instead of only searching the public web.

You can also completely disable web searches and only run internal data analysis.

Native chart output: Reports can directly generate HTML charts or info-graphics, eliminating the need to copy data into Excel for visualization. This serves as an intermediate product from plain text reports to visual analysis files, and you can further modify or refine these visuals later.

Collaborative research planning: The agent won’t run immediately; it will first produce a research plan. You can review, modify, and confirm it before execution. If you’re dissatisfied with the results, you can continue multi-turn conversations to ask questions or adjust the direction.

Currently, both agents are in Public Preview, accessible via Gemini API paid tiers. To try them, you can operate directly on Google AI Studio. Python SDK and JavaScript SDK are already supported, and those with free usage credits should not miss out.