Google on Thursday introduced a major upgrade to its AI research agent, Gemini Deep Research, calling it the company’s most advanced move so far into agentic artificial intelligence. The launch coincided with OpenAI’s release of GPT-5.2, further heating up the global race for AI leadership.
The upgraded Gemini Deep Research is now powered by Google’s flagship Gemini 3 Pro foundation model. According to Google, the new version is built to handle large-scale information synthesis, long and complex prompts, and advanced reasoning tasks with greater accuracy. The company says Gemini 3 Pro is its “most factual” model yet, specifically trained to reduce hallucinations, a key challenge in multi-step, autonomous AI systems.
Unlike earlier versions that primarily generated research reports, the new Gemini Deep Research is designed with developers in mind. Through Google’s newly launched Interactions API, developers can embed deep research-level AI capabilities directly into their own applications. This allows greater flexibility and control as AI agents increasingly become central to digital products and services.
Google says the tool is already being used for demanding workflows such as corporate due diligence, large-scale analysis, and drug toxicity safety research, where reliability and precision are critical.
The company also announced plans to integrate Gemini Deep Research across its wider ecosystem, including Google Search, Google Finance, the Gemini app, and NotebookLM. This move reflects Google’s vision of a future where users rely more on AI agents rather than traditional search queries to access and analyze information.
To support its performance claims, Google unveiled a new open-source benchmark called DeepSearchQA, designed to evaluate multi-step information-seeking and reasoning tasks. The agent was also tested on independent benchmarks such as Humanity’s Last Exam, a highly challenging general knowledge test, and BrowserComp, which measures browser-based agent performance.
Google said Gemini Deep Research outperformed competitors on its own benchmark and on Humanity’s Last Exam. However, OpenAI’s ChatGPT 5 Pro closely matched its results and slightly outperformed it on BrowserComp.
The competitive landscape shifted again later the same day when OpenAI released GPT-5.2, codenamed Garlic. OpenAI claims its new model exceeds Google’s performance across several benchmarks, including some developed internally. The near-simultaneous launches underscored the intense rivalry between the two companies, as both race to lead the next generation of advanced AI systems.
