Google Announces Deep Research Max: Autonomous Research Agents on Gemini 3.1 Pro
▶ ⏱️ Key moments
Google has launched Deep Research Max on Gemini 3.1 Pro: two agent configurations that split the difference between speed and thoroughness, with MCP support for private data and native chart generation built in.
Where this is going
If you’ve ever waited weeks for a research report that a well-trained intern could draft in days, this announcement is aimed squarely at you. Google is betting that autonomous research agents, not chatbots, are the next interface for knowledge work. Deep Research and Deep Research Max are two flavors of the same idea: let the machine do the legwork, then hand you a report you can actually use. Here’s what changed, what’s new, and where the tradeoffs land.
The story
Two agents, two philosophies
Google replaced the December Deep Research release with two distinct configurations. Deep Research (standard) is built for speed and interactivity: lower latency, lower cost, good enough for dashboards and near-real-time workflows. Deep Research Max cranks up test-time compute to iteratively reason, search, and refine its output, trading speed for depth.
The numbers Google reports:
| Benchmark | Deep Research Max |
|---|---|
| DeepSearchQA | 93.3% |
| HLE (Humanity’s Last Exam) | 54.6% |
Deep dive. The split reflects a structural tradeoff in AI agent design that’s been brewing for a while. As VentureBeat noted, standard Deep Research targets financial dashboards and real-time responses, while Max is built for asynchronous workflows like overnight due diligence reports ready by morning.
MCP: your private data, inside the agent
The biggest architectural shift is support for the Model Context Protocol (MCP). Deep Research can now query private databases, internal document repositories, and third-party data services without sensitive information leaving its source environment. Google is partnering with FactSet, S&P Global, and PitchBook to pipe their financial data streams through MCP servers.
This matters because it turns Deep Research from a web-only researcher into something that can reason over your proprietary data. The data stays where it is; the agent reaches in.
Charts, PDFs, and multimodal grounding
For the first time in the Gemini API, Deep Research natively generates charts and infographics within reports, using HTML or the Nano Banana format. The agent also accepts multimodal inputs for research grounding:
- PDFs
- CSVs
- Images
- Audio
- Video
The idea: don’t just search the web, search everything you have.
What early adopters are saying
The video features two partner testimonials worth noting. FactSet’s team emphasizes data reliability: “You can have innovation and the most advanced features, but if the data is not rock-solid, our customers will not use it.” Axiom, which predicts clinical trial outcomes, highlights the ability to pull information from deep inside complex documents: “Often what you need to know is on page 80 of a very long PDF.”
Both point to the same thing: the value isn’t in the model’s fluency, it’s in whether the output is grounded in something you can trust.
Availability
Deep dive. Both agents are available in public preview through paid tiers of the Gemini API, accessible via the Interactions API introduced in December 2025. Additional features include:
- Collaborative planning: review the research plan before execution
- Real-time streaming: watch intermediate steps as they happen
- Multimodal input grounding: feed PDFs, CSVs, images, audio, and video as context
The takeaway
Key points:
- Deep Research Max splits the agent model in two: fast/interactive vs. thorough/asynchronous, and that tradeoff is now explicit rather than hidden
- MCP support means the agent can reason over proprietary data without moving it, which is the real unlock for enterprise adoption
- Native chart generation and multimodal input grounding make the output more than just text
The research agent isn’t coming. It’s here. The question is whether enterprises will trust it enough to hand it the keys to their data.