Google Launches Deep Research and Max Agents on Gemini
Google unveiled two autonomous research agents, Deep Research and Deep Research Max, both powered by the Gemini 3.1 Pro model. Developers can access them now in public preview through the paid tiers of the Gemini API. A single API call starts a complete research process. For the first time, these agents draw information from the open web and private data sources to produce reports with full citations.
Google DeepMind, the division behind advanced AI systems like AlphaFold for protein folding and earlier Gemini models released starting in late 2023, developed these agents. Gemini represents Google's family of multimodal large language models capable of handling text, images, and more. The new agents aim to assist with intensive research duties that typically demand significant human effort.
Versions Tailored to Specific Needs
The regular Deep Research agent improves on a preview version Google shared in December. It delivers higher quality results at reduced latency and cost. This suits scenarios demanding fast replies, such as chat applications where users await instant feedback.
Deep Research Max takes a different approach. It focuses on comprehensive analysis rather than quick output. The agent employs extra processing time during testing to think, search, and refine its report. Google recommends it for background tasks that run asynchronously, for example, a scheduled job that generates a detailed due diligence report overnight for analysts.
Strong Results in Benchmarks
In tests conducted by Google, Deep Research Max excels in retrieving information and reasoning. It consults more sources than the prior version and identifies details that the earlier model overlooked. Google states these gains make it superior to its predecessor.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Comparisons with rivals show mixed clarity. The agent outperforms Google's previous iteration clearly. However, direct matches with OpenAI's GPT-5.4 or Anthropic's Opus 4.6 prove tricky due to differences in setup. OpenAI's GPT-5.4 handles web searches well but lacks tuning for deep research. OpenAI offers a dedicated DR agent, updated to GPT-5.2 in February, not GPT-5.4. OpenAI's top search option, GPT-5.4 Pro, scored 89.3 percent on the BrowseComp benchmark for agentic search, compared to 82.7 percent for GPT-5.4.
Anthropic reports 84 percent for Opus 4.6 on BrowseComp, achieved without reasoning enabled, as it performed better that way than under the high reasoning settings Google applied. Variations stem from evaluation methods, API access versus proprietary tools, and other factors. Google presents its figures, but users should approach them carefully given the lack of full transparency.
Key Enhancements Including MCP Integration
Support for the Model Context Protocol (MCP) marks a major upgrade. Developers connect the agents to custom data sources, such as financial databases or market feeds. The agents accept any tool specifications, evolving from web-only searchers into versatile systems that query specialized repositories.
The Gemini API now allows agents to create charts and infographics natively within reports. Outputs appear in HTML or "Nano Banana" format for straightforward visual display of intricate data.
Additional capabilities cover collaborative planning, where users can adjust the agent's search strategy beforehand. Inputs support multiple formats: PDFs, CSVs, images, audio, and video. Real-time updates show progress during execution. Developers may disable web access to restrict operations to internal data alone.
These agents operate on infrastructure supporting research tools in Google's products, including the Gemini app, NotebookLM, Google Search, and Google Finance. Builders use the Interactions API for custom workflows. Rollout to startups and enterprises follows via Google Cloud.