Google on Thursday launched a powerful new version of its AI research agent, Gemini Deep Research, marking its most advanced step yet into agentic AI.

The update arrived on the same day OpenAI released GPT-5.2, intensifying the industry’s race for dominance. The tool is now built on Google’s flagship Gemini 3 Pro foundation model.

Google’s deepest research agent yet

Powered by Gemini 3 Pro

The reimagined Gemini Deep Research agent is designed to handle massive information synthesis, long prompts and complex reasoning tasks. Google says the upgrade benefits from Gemini 3 Pro’s status as its “most factual” model, engineered to minimize hallucinations — a critical requirement for deep, multi-step reasoning.

Designed for developers 

Unlike earlier versions, the new Deep Research tool is not limited to generating research reports.
Through Google’s new Interactions API, developers can embed Deep Research–level capabilities directly into their applications, giving them far more control as AI agents become central to future products.

Google says customers already use the agent for extensive workflows such as due diligence and drug toxicity safety analysis.

Integration across Google’s ecosystem

The company also announced plans to integrate this agent into products including:

• Google Search

• Google Finance

• Gemini App

• NotebookLM

The move signals a shift toward a future where individuals may rely on AI agents—not traditional search boxes—to find information.

AI hallucinations remain one of the biggest hurdles in agentic systems that make numerous autonomous choices.
Google says Gemini 3 Pro is specifically trained to reduce these errors in long-running, high-stakes tasks.

Google creates new benchmark

To back its claims, Google introduced a new benchmark called DeepSearchQA, designed to test multi-step, information-seeking tasks. The benchmark has been open-sourced.

The agent was also evaluated on independent tests such as:

• Humanity’s Last Exam — a notoriously difficult general-knowledge benchmark

• BrowserComp — focused on browser-based agent tasks

Google’s Deep Research outperformed rivals on its own benchmark and Humanity’s Last Exam. However, OpenAI’s ChatGPT 5 Pro came surprisingly close across categories and even edged ahead on BrowserComp.

OpenAI responds with GPT-5.2

The competitive picture shifted almost instantly when OpenAI rolled out its long-awaited GPT-5.2, codenamed Garlic, later the same day.
OpenAI claims the model surpasses Google’s on multiple benchmarks, including those it developed itself.

The timing raised eyebrows — with both companies appearing eager to outshine the other on a day filled with major AI announcements.