Google LLC today released a new version of Gemini Deep Research, an artificial intelligence agent designed to automate complex tasks such as crafting financial reports.

The company first introduced the tool last December. The initial version used Gemini 1.5 Pro, which at the time was Google’s flagship large language model. The new version that debuted today is based on Gemini 3 Pro, a significantly more capable model released last month.

One of the areas where Gemini 3 Pro performs better than its predecessors is visual reasoning. According to Google, it can perform tasks such as planning the travel paths of a warehouse robot. When the LLM is applied to document processing use cases, it can extract information from handwritten text, charts and mathematical notation.

The new release of Gemini Deep Research uses Gemini 3 Pro’s visual reasoning features to automate data retrieval tasks. Users can upload documents and have the agent scan them to find a specific piece of information. Alternatively, Gemini Deep Research can be instructed to condense the documents into a report or enrich them with information from the web.

Google says that the agent’s new release introduces significantly improved web search capabilities. “Deep Research iteratively plans its investigation – it formulates queries, reads results, identifies knowledge gaps, and searches again,” Google DeepMind product manager Lukas Haas and group product manager Shrestha Basu Mallick explained in a blog post.

Gemini Deep Research is available through a new application programming interface called Interactions API that debuted in conjunction. It enables developers to access the agent and the Gemini model series through a single access point. In the future, Google will add more prepackaged agents to interactions API along with support for custom agent development.

Besides providing centralized access to multiple AI offerings, the API also eases certain programing tasks. It automates some of the work involved in managing the data that users upload to an AI application for processing. Additionally, there’s an MCP tool for connecting AI models to third-party systems.

Google evaluated Gemini Deep Research’s capabilities using two benchmarks called HLE and DeepSearchQA. According to the company, it achieved record performance on both tests.

HLE is a particularly difficult AI benchmark that comprises over 2,500 questions. More than half of the questions relate to math, physics and programming. Google says that Gemini Deep Research solved 46.4% of the problems in HLE correctly.

The other benchmark that Google used in the evaluation, DeepSearchQA, is an internally-developed dataset it open-sourced today. It comprises 900 multi-step tasks in which each step “depends on prior analysis.” Google says that the benchmark measures the precision and comprehensiveness of AI models’ research. 

In addition to measuring LLM output quality, DeepSearchQA can help researchers find ways of refining their algorithms. It “serves as a diagnostic tool for the benefits of ‘thinking time,’” Haas and Basu Mallick wrote. “In our internal evaluations, we observed significant performance gains when allowing the agent to perform more searches and reasoning steps which we plan to explore in future releases.”

Image: Google