Just before OpenAI announced its GPT-5.2, codenamed "Garlic," Google took the lead by launching its newly upgraded AI research agent—Gemini Deep Research. Built on the latest Gemini 3 Pro large model, this agent not only generates high-quality research reports but more importantly, it is now open to developers through a brand-new Interactions API, allowing Google's advanced research capabilities to be embedded into third-party applications.
This marks Google's accelerated move toward an "agent-based AI" era—a future where humans no longer search for information themselves, but instead have AI agents handle complex information tasks. Gemini Deep Research is designed to handle massive amounts of information and ultra-long context prompts, efficiently integrating complex data flows, and is suitable for high-demand scenarios such as due diligence and drug toxicity safety assessments. Google stated that this tool will soon be integrated into Google Search, Google Finance, the Gemini App, and the widely popular NotebookLM among academic circles.
To address the "hallucination" issue commonly seen in agent-based AI during long-term reasoning tasks—where large models fabricate content out of thin air—Gemini 3 Pro has been specifically optimized for factual accuracy. In multi-step, long-running autonomous decision-making chains, even a single error in reasoning could invalidate the entire output, making model reliability crucial.
To validate its technical advantages, Google also launched a new benchmark test called DeepSearchQA, specifically designed to evaluate AI agents' performance in complex, multi-hop information retrieval tasks, and has already open-sourced the evaluation set. Additionally, Gemini Deep Research made appearances in two external authoritative tests: one is the notorious "Humanity’s Last Exam," known for its challenging and obscure questions, and the other is BrowserComp, focused on browser automation tasks. The results showed that Google's agent led in the first two tests but lagged slightly behind OpenAI's ChatGPT 5 Pro in BrowserComp.
Sarcasm aside, these comparison data almost immediately lost their relevance—because just a day after the release, OpenAI officially launched GPT-5.2, claiming it surpassed competitors in multiple tests, including its own benchmarks, especially highlighting its victory over Google. This precise timing of the launch highlights the fierce competition between the two AI giants in the agent intelligence arena: one attempts to define the next generation of AI assistants with deep research capabilities, while the other counters with more general reasoning performance. The real AI agent war, perhaps, has just begun.
