Google unveils Gemini Deep Research agent to rival OpenAI's 'Garlic'

Google on Thursday unveiled a ‘reimagined’ version of its research agent, Gemini Deep Research, based on its much-ballyhooed state-of-the-art foundation model, Gemini 3 Pro.

While producing a research report, it now lets developers embed Google’s SOTA-model (state-of-the-art) research capabilities into their own apps. That capability is made achievable through Google’s new Interactions API, which is designed to provide devs more control in the coming agentic AI era.

Google’s new Gemini Deep Research tool is an advanced agent designed to synthesize large volumes of information and manage extensive context dumps within the prompt. According to Google, this tool is currently used by customers for a variety of tasks, from conducting due diligence to researching drug toxicity.

Furthermore, Google has announced plans to soon integrate this deep research agent into several of its services, including Google Search, Google Finance, the Gemini App, and the widely used NotebookLM. This move signifies a broader trend toward a future in which AI agents, rather than humans, will handle information retrieval and search.

Google says that Deep Research benefits from Gemini 3 Pro’s status as its “most factual” model, which is trained to reduce hallucinations during tough tasks.

AI hallucinations, in which Large Language Models (LLMs) fabricate information, pose a particularly significant problem for complex, long-running agentic tasks. These tasks involve numerous autonomous decisions over extended periods (minutes, hours, or longer). Since the probability of a hallucinated choice invalidating the entire result increases with the number of decisions an LLM must make, this issue is especially critical.

To prove its improvement claims, the tech giant has also created yet another benchmark named DeepSearchQA, which is intended to test agents on complex, multi-step information-seeking tasks. Google has open-sourced this benchmark.

The tech company recently tested its new AI agent against several benchmarks, including Deep Research on Humanity’s Last Exam, which is a unique independent measure of general knowledge with exceptionally niche tasks, and BrowserComp, which evaluates browser-based tasks requiring agency.

As expected, Google’s new agent outperformed the competition on both its own benchmark and Humanity’s Last Exam. However, OpenAI’s ChatGPT 5 Pro performed surprisingly well, coming in a close second overall and slightly surpassing Google on the BrowserComp test.

Yet, those benchmark comparisons became outdated almost immediately after Google released them, as OpenAI launched its highly anticipated GPT 5.2, codenamed Garlic, on the same day. OpenAI claims that its latest model outperforms its rivals, particularly Google, across a variety of standard benchmarks, including its own.

Perhaps the most interesting aspect of this announcement was the timing. Knowing that the world was awaiting the release of Garlic, Google strategically dropped major AI news of its own to share the spotlight.