Google’s Deepest AI Agent Debuts Amid OpenAI’s GPT-5.2 Launch
▼ Summary
– Google released a new version of its Gemini Deep Research agent, rebuilt on its Gemini 3 Pro model, which is designed to synthesize large amounts of information.
– The tool now allows developers to embed its research capabilities into their own apps via a new Interactions API, extending its use beyond just generating reports.
– Google plans to integrate this deep research agent into services like Google Search, Finance, the Gemini App, and NotebookLM, moving toward AI agents performing searches for users.
– The agent leverages Gemini 3 Pro’s factuality to minimize AI hallucinations, which is critical for long, complex reasoning tasks where errors can invalidate the entire output.
– Google tested the agent on new benchmarks, where it performed well, but the competitive landscape shifted immediately as OpenAI launched its rival GPT 5.2 model on the same day.
Google has introduced a significantly upgraded version of its Gemini Deep Research agent, built upon the advanced Gemini 3 Pro foundation model. This move represents a strategic push to embed sophisticated AI research capabilities directly into third-party applications and Google’s own ecosystem. The enhanced agent is engineered to process vast amounts of information and manage extensive contextual prompts, positioning it as a powerful tool for complex analytical work. Its primary function extends beyond generating reports to enabling developers to integrate its deep research functionalities into their own software via a new Interactions API. This development signals a shift towards an “agentic” future where AI handles intricate information-seeking tasks that users would traditionally perform themselves.
The applications for this technology are substantial, with current use cases spanning critical fields like financial due diligence and pharmaceutical safety research. Google has announced plans to weave this deep research agent into several of its flagship services, including Google Search, Google Finance, the Gemini App, and NotebookLM. This integration foreshadows a paradigm where AI assistants proactively conduct comprehensive research on behalf of users, fundamentally changing how we interact with information.
A central challenge for AI agents engaged in long, multi-step reasoning tasks is the risk of “hallucinations”, instances where the model generates plausible but incorrect information. Google emphasizes that Gemini 3 Pro is its “most factual” model to date, specifically trained to minimize these errors during complex operations. The integrity of an agent’s entire output can be compromised by even a single flawed decision, making this focus on accuracy crucial for reliable deep research.
To substantiate its performance claims, Google has released a new open-source benchmark named DeepSearchQA, designed to evaluate agents on complicated, multi-step information retrieval. The company also tested its agent on independent benchmarks like the intriguingly titled Humanity’s Last Exam and BrowserComp. While Google’s new agent performed strongly, notably leading on its own benchmark and Humanity’s Last Exam, OpenAI’s ChatGPT 5 Pro proved to be a formidable competitor, placing a close second in most tests and even slightly outperforming Google on the BrowserComp benchmark.
However, the competitive landscape shifted almost immediately. On the very same day as Google’s announcement, OpenAI launched its latest model, GPT 5.2, internally codenamed “Garlic.” OpenAI asserts that this new iteration surpasses rival models, including Google’s, across a standard suite of industry benchmarks. The synchronized timing of these releases highlights the intense, rapid-fire competition dominating the AI industry, as leading companies vie to demonstrate superior capabilities and capture developer mindshare.
(Source: TechCrunch)
