On Thursday, Google unveiled a "reimagined" iteration of its research agent, Gemini Deep Research, built upon its advanced foundation model, Gemini 3 Pro. This updated agent goes beyond generating research reports—though it retains that function—by enabling developers to integrate Google's sophisticated research capabilities directly into their own applications.

This new functionality is delivered through Google's freshly introduced Interactions API, crafted to provide developers with enhanced control as the industry moves further into an era dominated by agentic AI. The tool itself is designed as an agent capable of synthesizing vast amounts of information and processing extensive context within a single prompt.

According to Google, customers are already utilizing it for a diverse range of tasks, from conducting due diligence to researching drug toxicity safety. The company also announced plans to soon incorporate this deep research agent into several of its services, including Google Search, Google Finance, the Gemini App, and its popular NotebookLM. This integration marks another stride toward a future where AI agents, rather than humans, perform the bulk of information retrieval and analysis.

Google emphasizes that Deep Research leverages Gemini 3 Pro's distinction as its "most factual" model, specifically trained to reduce hallucinations during complex tasks. This is particularly critical for extended, deep-reasoning agentic work, where a series of autonomous decisions are made over prolonged periods. In such scenarios, even a single fabricated choice by the model could compromise the entire output.

To substantiate its progress, Google has introduced a new benchmark named DeepSearchQA, aimed at evaluating agents on complex, multi-step information-seeking tasks. This benchmark has been open-sourced. Additionally, Google tested Deep Research on two other benchmarks: the independently created Humanity’s Last Exam, which tests broad knowledge with highly niche tasks, and BrowserComp, a benchmark for browser-based agentic tasks.

As anticipated, Google's new agent outperformed competitors on its own benchmark and on Humanity’s Last Exam. However, OpenAI’s ChatGPT 5 Pro proved to be a surprisingly close second across these evaluations and even slightly edged out Google on the BrowserComp benchmark.

Notably, these comparative results were quickly overshadowed. On the very same day, OpenAI launched its highly anticipated GPT-5.2 model, codenamed Garlic. OpenAI states that this newest model surpasses its rivals, particularly Google, across a suite of standard benchmarks, including one of its own creation.

The timing of Google's announcement is especially interesting, arriving just as the industry was anticipating the release of OpenAI's Garlic, suggesting a strategic move in the ongoing AI landscape.