Google introduced a redesigned Gemini Deep Research agent on Thursday, built upon its Gemini 3 Pro foundation model. This updated agent allows developers to integrate Google's advanced research capabilities directly into their proprietary applications via a new Interactions API. Google specified that the agent can synthesize extensive datasets and manage substantial contextual inputs, supporting diverse applications from corporate due diligence processes to pharmaceutical drug toxicity safety research.
The introduction of the Interactions API is positioned to provide developers with enhanced control in an increasingly agentic artificial intelligence ecosystem. Google emphasized that the Deep Research agent benefits from Gemini 3 Pro's design as its "most factual" model, trained to mitigate "hallucinations"—instances where the AI generates incorrect information—during complex, multi-step tasks. This reduction in factual errors is deemed crucial for autonomous agentic operations, particularly those involving extended reasoning over minutes or hours, where a single inaccuracy could compromise the entire output.
Looking ahead, Google plans to integrate the new deep research agent into several of its existing services. These integrations are slated for Google Search, Google Finance, the Gemini App, and NotebookLM, indicating a strategic direction towards embedding advanced AI research functionalities across its product portfolio.
Concurrently with the agent's release, Google unveiled DeepSearchQA, an open-sourced benchmark developed to evaluate AI agents on intricate, multi-step information retrieval tasks. Tests conducted on DeepSearchQA, alongside independent benchmarks such as Humanity’s Last Exam and BrowserComp, indicated the new Google agent's strong performance, leading in the company's proprietary benchmark and Humanity's Last Exam. However, OpenAI's ChatGPT 5 Pro closely followed in performance and marginally outperformed Google's agent on BrowserComp. This release coincided with OpenAI's launch of GPT 5.2, codenamed Garlic, which OpenAI asserts surpasses competing models, including Google's, across a suite of standard benchmarks. This simultaneous timing highlights the accelerating competition in the artificial intelligence development sector.
