Google released a "reimagined" version of Gemini Deep Research based on its most advanced foundation model, Gemini 3 Pro, on Thursday. This new intelligent agent not only retains the core function of generating research reports but also, more importantly, allows developers to embed Google's SATA model research capabilities into their own applications for the first time through a new interactive API, giving developers greater control in the age of agent AI.
Functional Upgrades and Wide Applications
This intelligent agent is capable of integrating massive information and processing large context data within prompts. According to Google, the range of tasks performed by clients using this tool is extensive, covering everything from corporate due diligence to drug toxicity safety research, demonstrating its practical value in professional fields.
Google also announced that this deep research agent will be integrated into several of its core services, including Google Search, Google Finance, the Gemini app, and the popular NotebookLM. This move marks Google's step toward a future vision where AI agents replace humans in performing search tasks.
Technical Breakthroughs: Combating AI Hallucinations
Google emphasized that Deep Research benefits from Gemini 3 Pro's status as its "most accurate" model. This model has been specifically trained to minimize hallucinations when performing complex tasks. AI hallucinations—where large language models fabricate information at will—are particularly critical for long-running deep reasoning tasks that require many autonomous decisions over minutes, hours, or even longer. Even a single hallucinatory choice can render the entire output invalid.
Performance in Benchmark Tests
To demonstrate technical advancements, Google created a new benchmark test called DeepSearchQA and has made it open-source, specifically testing agents' performance in complex multi-step information retrieval tasks. Google also tested the "Humanity's Last Exam Deep Research" benchmark, which tests general knowledge, and the BrowserComp, a browser-based agent task benchmark.
Test results showed that Google's new agent outperformed competitors in its self-developed benchmarks and the Humanity benchmark. However, OpenAI's ChatGPT5Pro performed surprisingly well, coming closely behind in all tests and slightly surpassing in the BrowserComp test.
Intensifying Market Competition
However, these comparisons became almost obsolete almost immediately. On the same day, OpenAI released its highly anticipated GPT5.2 (codename Garlic). OpenAI stated that its latest model outperformed competitors, especially Google, in a series of standard benchmark tests, including those developed by OpenAI itself.
The most notable aspect of this announcement was its timing. Google knew the world was eagerly waiting for the release of Garlic and simultaneously announced this new AI update, highlighting the intense competition between tech giants in the AI field.
