Google announced a significantly enhanced version of its Gemini Deep Research agent, now accessible through the Interactions API. The update provides developers with direct integration of Google’s most advanced autonomous research technology, enabling long-running investigation, synthesis, and reporting across large information sets. The announcement arrived on the same day OpenAI unveiled GPT-5.2, underscoring the competitive pace of progress in advanced AI systems.
The Deep Research agent is built on Gemini 3 Pro, which Google describes as its most factual model to date. The agent has been trained to reduce hallucinations and improve reliability in complex research scenarios. To support this, Google scaled multi-step reinforcement learning for search, allowing the agent to move through intricate information sources, evaluate context, and refine its queries with higher precision.
The system conducts research iteratively. It generates queries, reviews results, identifies knowledge gaps, and continues searching until the task is completed. Google said this release includes substantially improved web navigation, enabling the agent to explore deeper sections of websites for specific data points.
Launch of DeepSearchQA Benchmark
To evaluate agents operating in real-world research conditions, Google open-sourced DeepSearchQA, a benchmark designed for multi-step, causal reasoning across the web. The dataset spans 900 tasks across 17 fields, with each task requiring sequential, dependency-based analysis. Unlike traditional fact-retrieval benchmarks, DeepSearchQA focuses on comprehensiveness and full answer set generation.
Google said the benchmark also demonstrates the benefits of extended “thinking time” during evaluations. Internal tests showed that allowing more reasoning steps produced measurable quality improvements. The company is releasing the dataset, leaderboard, and technical report to support broader agent research.
Early Use Cases in Technical and Financial Fields
According to Google, the Gemini Deep Research agent is already being used in industries that require high-precision research workflows. Financial firms reported that the agent can automate the initial stages of due diligence, consolidating market data, competitive signals, and compliance information from public and proprietary sources. Early users said the tool has shortened research cycles from days to hours.
In biotech, Axiom Bio used Gemini Deep Research to analyze biomedical literature at a level of depth the company described as previously achievable only through expert manual review. The agent’s ability to surface granular evidence was cited as a foundation for more advanced AI systems that support drug safety and discovery.
Developer Capabilities Through the Interactions API
Developers can now use the Interactions API to embed Gemini Deep Research into applications for automated analysis and report generation. The agent can review PDFs, spreadsheets, and documents and combine them with public web data through the File Search Tool. It handles large context inputs and supports structured outputs, including JSON schemas.
Additional features include controllable report formatting, detailed citations, and integration flexibility for downstream systems. Google said future releases will incorporate native chart generation, expanded Model Context Protocol connectivity, and enterprise availability through Vertex AI.
