Google has introduced a major upgrade to its AI research capabilities with the launch of a reworked Gemini Deep Research agent, powered by its most advanced foundation model, Gemini 3 Pro. The announcement landed on the same day OpenAI released GPT-5.2, underscoring just how intense the competition has become at the top end of the AI landscape.
The updated Gemini Deep Research is no longer limited to generating long-form research reports. Google has expanded the agent into a flexible research engine that developers can embed directly into their own applications.
This expansion is enabled through Google’s newly announced Interactions API, which is designed to give developers greater control as AI systems move toward more autonomous, agent-driven workflows.
Built for Large-Scale, Complex Research Tasks
Google describes Gemini Deep Research as an agent capable of synthesizing vast amounts of information while handling extremely large prompts and extended context windows. According to the company, customers are already using it for demanding tasks such as corporate due diligence, safety analysis, and pharmaceutical toxicity research.
The agent’s strength comes from its ability to manage long-running reasoning processes, where multiple decisions must be made sequentially without losing accuracy or coherence.
Google has also confirmed plans to integrate the new Deep Research agent across several of its core products, including Google Search, Google Finance, the Gemini app, and NotebookLM. This signals a broader strategy aimed at a future where people rely less on traditional search and more on AI agents that retrieve, analyze, and synthesize information on their behalf.
Designed to Reduce Hallucinations in Agentic Workflows
One of Google’s key claims is that Gemini Deep Research benefits from Gemini 3 Pro being its most factually reliable model to date. The model has been trained specifically to reduce hallucinations, a critical requirement for deep, multi-step reasoning tasks.
Hallucinations pose a serious risk in agent-based systems. When an AI agent makes dozens or even hundreds of autonomous decisions over extended periods, a single fabricated assumption can compromise the entire result. As agentic AI becomes more common, factual consistency becomes just as important as raw intelligence.
Google positions Gemini Deep Research as a response to this challenge, emphasizing reliability alongside scale and reasoning depth.
New Benchmarks to Prove Progress
To support its claims, Google introduced a new benchmark called DeepSearchQA, designed to evaluate how well AI agents perform on complex, multi-step information retrieval tasks. The company has made this benchmark open source, allowing the broader research community to test and compare results.
In addition to its own benchmark, Google also evaluated Gemini Deep Research on established tests such as Humanity’s Last Exam, which focuses on highly specialized general knowledge, and BrowserComp, a benchmark that measures how well agents handle browser-tasks.
On its own DeepSearchQA benchmark and Humanity’s Last Exam, Google’s agent outperformed competitors. However, OpenAI’s ChatGPT 5 Pro ranked closely behind and edged ahead of Google on BrowserComp, highlighting how narrow the performance gap has become between leading models.
A Release Overshadowed by GPT-5.2
Despite Google’s strong benchmark showing, the comparisons became outdated almost immediately. On the same day, OpenAI released GPT-5.2, internally codenamed Garlic.
OpenAI claims that GPT-5.2 outperforms rival models across a wide range of standard benchmarks, including several developed in-house. The timing of both announcements was hard to ignore, especially given that the industry had been anticipating OpenAI’s release for weeks.
Google’s decision to announce its deepest research agent on the same day suggests a deliberate effort to remain visible and competitive in a fast-moving AI arms race.
A Signal of the Industry’s Direction
The near-simultaneous launches highlight a broader shift in the AI industry. The focus is moving away from simple chatbots and toward agentic systems that can reason over long time horizons, interact with tools, and make autonomous decisions with minimal human oversight.
In this new phase, success will depend not only on benchmark scores, but on reliability, integration, and the ability to operate safely at scale.
For both Google and OpenAI, these releases are less about a single model update and more about staking a claim in what comes next.