Table of Contents
- What Google Actually Shipped
- Deep Research vs Deep Research Max
- The MCP Adoption Is the Story
- What the 93% Benchmark Actually Means
- How to Use Each Agent
- What This Signals for the Rest of 2026
- FAQ
Two years ago, Anthropic published the Model Context Protocol specification. The pitch was modest: a standard way for any AI agent to connect to any data source or tool. The bet was that if enough labs adopted MCP, the agent ecosystem would consolidate around it the way the web consolidated around HTTP. The bet looked plausible. It did not look settled.
On April 21, 2026, Google shipped Deep Research and Deep Research Max — two new autonomous research agents built on Gemini 3.1 Pro, designed to connect open-web search with proprietary enterprise data and produce reports with native charts. Most coverage led with the 93.3% benchmark score on DeepSearchQA. Useful number. The structural detail that matters more is buried two paragraphs down: MCP support ships native in both products.
That sentence is the moment the MCP standard war ended. Anthropic invented the protocol. OpenAI adopted it for Codex and ChatGPT. Cursor 3 built around it. Google now ships it in flagship Gemini products. The four largest AI labs in the world all speak the same agent-tooling protocol. Whatever Google’s competitive instinct said about not adopting a competitor’s spec, the product team’s instinct said the cost of fragmentation was higher.
What Google Actually Shipped
Deep Research and Deep Research Max are autonomous research agents that take a natural-language query, run multi-step web research, fuse results with any proprietary data sources you connect, and produce a structured report with embedded charts. The agents replace Google’s December 2025 preview Deep Research agent, which was a research demonstration without production-grade reliability.
Both products are powered by Gemini 3.1 Pro. Both ship with native MCP support for custom data integration. Both produce native charts and infographics inline in their reports rather than describing data in prose. Both are available in public preview through paid tiers of the Gemini API via the Interactions API surface Google introduced in late 2025.
The shipped feature set is meaningfully ahead of what was available three months ago. Two-step research with citations is now standard across frontier labs. Multi-step research with native MCP-connected data sources and inline visualizations is, as of April 21, available from exactly one vendor at production-grade reliability.
Deep Research vs Deep Research Max
Google’s tier structure encodes a real tradeoff. Same model, different inference posture.
Deep Research is the standard tier. Optimized for low latency. Designed to be embedded inside interactive surfaces where a user is waiting for an answer. Typical query completion time is seconds to a few minutes. Best for chatbot integrations, support-tool augmentation, research summaries inside live workflows.
Deep Research Max is the extended-compute tier. Uses substantially more reasoning time per query, iterates against its own findings, refines reports through multiple passes. Typical query completion time is minutes to hours. Designed for asynchronous, background tasks — the overnight due-diligence brief, the comprehensive industry scan, the deep-dive whose value comes from thoroughness rather than speed.
The split is not about model quality. The model is the same. The split is about how much compute the user is willing to spend per query. For roughly 80% of research tasks, the standard tier is enough. For the 20% where comprehensiveness materially changes the answer — typically high-stakes, long-tail, decision-relevant research — Max earns its compute budget.
The MCP Adoption Is the Story
Pull together the last twelve months of agent-stack standardization and the pattern is hard to miss.
November 2024. Anthropic publishes the Model Context Protocol as an open spec. Initial adoption is Anthropic-only.
Q1 2025. The first wave of third-party MCP servers ship. Linear, Notion, GitHub, and others publish reference connectors. The ecosystem starts to look like a real standard rather than an Anthropic-internal abstraction.
December 2025. OpenAI adopts MCP for the Codex CLI and the ChatGPT desktop app. The spec now spans two of the three largest labs. GPT-5.5’s launch coverage noted MCP as a core dependency.
Q1 2026. The Realtime API gets MCP support. Voice agents now share the same tool-connection protocol as coding agents.
April 2026. Cursor 3 ships with MCP at the center of its Agents Window architecture. Every agent in Cursor 3 speaks MCP natively.
April 21, 2026. Google adopts MCP in Deep Research and Deep Research Max. The third major lab is in.
That is what a settled standard looks like. The remaining holdouts in the agent space — Meta, xAI, the major Chinese labs — are now adopting MCP for compatibility, not building competing protocols. The fragmentation risk for the agent ecosystem dropped meaningfully between Q4 2025 and Q2 2026, and Deep Research is the punctuation mark.
This matters because every infrastructure standard inflection point compounds. Once MCP is the agent-tooling protocol the way HTTP is the document protocol, every connector built for one lab works against every other. Every enterprise that builds a custom MCP server for their internal data plugs into Anthropic, OpenAI, Google, Cursor, and the rest without integration work. The cost curve for agent deployments drops accordingly.
What the 93% Benchmark Actually Means
Deep Research Max scored 93.3% on DeepSearchQA, the standard benchmark for autonomous research agents. That is a real number. It is also a number that needs context.
DeepSearchQA measures whether an agent can answer a multi-step research question correctly when given web access and unlimited reasoning time. Previous frontier scores were in the 70s. 93.3% is a substantial improvement, but the benchmark is constructed to measure correctness of factual research, not depth of synthesis or quality of presentation. A model can score 93% and still produce reports that are correct but not useful, or useful but not actionable.
What the benchmark does tell you reliably: Deep Research Max is meaningfully more accurate than its predecessor at the kind of research task where the answer is specific and verifiable. Industry deep-dives, due diligence on a specific company, technical research with a defined question — these get reliably better. Open-ended exploratory research, where the value comes from framing the right question rather than finding the right answer, depends more on prompt quality than benchmark score.
How to Use Each Agent
Both agents are available through the Gemini API. Set the model parameter to gemini-deep-research for the standard tier or gemini-deep-research-max for the extended-compute tier. The Interactions API surface handles the multi-turn research flow, which is different from a standard chat completion call — you submit a research goal, receive intermediate status updates, and pull the final report when complete.
Connect proprietary data via MCP. Point the agent at any MCP-compliant server, including custom servers built for internal databases, document stores, or specialized data providers. The same MCP servers you may already have running with Claude or Cursor will work with Deep Research without modification.
Do this first: take a research task that currently lives on someone’s plate as a multi-hour ad-hoc workflow — competitive intelligence, market sizing, technical due diligence — and run it through Deep Research Max against your normal data sources. Compare the output to what the human produced. The honest assessment is rarely “as good.” It is often “70% as good in 3% of the time,” and that ratio is what changes the staffing math.
What This Signals for the Rest of 2026
Two predictions worth tracking through Q3.
Every major lab will ship a Deep Research equivalent by Google I/O 2026. Anthropic already has Claude Research. OpenAI has Deep Research (the original, predating Google’s). Cursor 3 has /best-of-n and Agents Window equivalents. The remaining gaps are xAI and the Chinese labs. Expect both to ship comparable products within 90 days. The competitive pressure point is not whether you can do autonomous research — it is whether you can do it in your customer’s existing data context, which is why MCP adoption is the strategic pivot point.
The next infrastructure standardization fight is on agent identity and authorization. MCP solved the tooling connection question. The remaining open question is how agents prove who they are and what they are authorized to do across organizational boundaries. Expect proposed standards from at least two labs by the end of summer. Expect adoption to follow a faster curve than MCP did because the labs now have shared experience that standardization is the cooperative move.
Deep Research and Deep Research Max are good products. The benchmark score is real. The MCP adoption is the structural signal underneath the product launch — and structural signals compound longer than benchmark scores do.
FAQ
What is the difference between Deep Research and Deep Research Max?
Same underlying model (Gemini 3.1 Pro). Different inference posture. Deep Research is optimized for low-latency interactive use cases — typical completion in seconds to minutes. Deep Research Max uses extended reasoning time for asynchronous, comprehensive tasks — typical completion in minutes to hours. Use Deep Research inside live chatbot flows. Use Max for overnight due diligence.
Is MCP support new in Deep Research?
Yes. Google’s December 2025 preview Deep Research agent did not include MCP support. The April 21, 2026 launch is the first production Google AI surface with native MCP integration. This is the moment Google formally aligned with the agent-tooling standard Anthropic introduced and OpenAI adopted in late 2025.
How does Deep Research compare to Claude Research or OpenAI’s Deep Research?
All three are autonomous research agents on the same general architecture. Google’s 93.3% DeepSearchQA score leads the published benchmarks at launch. In practice, choose based on which model family your stack is already standardized on and which proprietary data connectors you need — MCP support makes cross-vendor portability higher than at any prior point.
What does the 93.3% DeepSearchQA score actually mean?
DeepSearchQA tests whether an agent can correctly answer multi-step research questions with web access and unlimited reasoning time. 93.3% is the highest score on the benchmark to date. It measures factual research accuracy, not synthesis quality or presentation. A high score means the agent is reliable on questions with verifiable answers; it does not directly measure how useful the resulting report is.
What does it cost to run Deep Research Max?
Pricing is structured around the Gemini API’s standard tiers with additional compute-time billing for extended reasoning. As of launch, Google has not published a per-query cost figure because the cost varies significantly with research depth. Plan for Max queries to cost noticeably more than standard Deep Research queries; reserve Max for tasks where comprehensiveness materially changes the answer.
Can I run Deep Research against my own internal data sources?
Yes, through MCP. Point either agent at any MCP-compliant server. Existing MCP servers built for other labs (Anthropic, OpenAI tooling) work without modification. This is the practical payoff of the standard adoption — the same connectors plug into the entire frontier-lab ecosystem.