They pause before an answer to ponder.
Introducing Gemini 2.5, our most intelligent AI model.
Our first release, an experimental version of 2.5 Pro, unlocks state-of-the-art performance in math and science.
Google (@Google) March 25, 2025
The first in the series is the experimental version of Gemini 2.5 Pro. The multimodal reasoning neural network "is the most intelligent to date." It is available on the Google AI Studio developer platform and in the Gemini app for Advanced subscription holders for $20 per month.
"In Gemini 2.5, we have achieved a new level of performance by combining a vastly improved base model with advanced post-learning," the corporation said in a blog post.
The company added that in the future, all of its models will have the ability to reason.
In the Aider Polyglot test, which tracks programming skills, Gemini 2.5 Pro scored 68.6%, beating leading solutions from OpenAI, Anthropic and DeepSeek.
In the SWE-bench Verified benchmark, which measures the abilities of software developers, Google's solution lost to Claude 3.7 Sonnet, 63.8% versus 70.3%.
In the multimodal test Humanity's Last Exam, which consists of a thousand questions in math, humanities and science, Gemini 2.5 Pro scored 18.8%, surpassing the results of its competitors;
Gemini 2.5 Pro has a contextual window of 1 million tokens - the model is capable of ingesting approximately 750,000 words at a time. There are plans to double the figure in the future.
"Reasoning" AI needs more time and computational resources to provide an accurate and extensive answer. The first "thinking" neural network showed was OpenAI - it released o1 in September 2024. Other AI developers have gotten busy creating their own solutions. Among them are DeepSeek, Anthropic, Alibaba and others.
To recap, in March, Google announced a number of AI products and features for the healthcare sector.
The company previously trained its AI assistant Gemini to analyze and take into account a user's search history to provide a more relevant answer.
Microsoft announces two AI agents for 'deep research'
Microsoft will add two "reasoning" AI agents to Microsoft 365 Copilot to analyze large amounts of information and provide highly qualified expertise - Researcher and Analyst.
Researcher is designed to help with the multi-step analysis process. The tool combines OpenAI's Deep Research mode for deep research with the "advanced search capabilities" of Microsoft 365 Copilot.
With Researcher, a user can develop a detailed go-to-market strategy, the company claims. The service will take into account business specifics and competitor information from the Internet, identify opportunities for a new product or create a quarterly report detailing performance.
Researcher is able to leverage information from other AI agents, tools and applications like Confluence, ServiceNow and Salesforce.
Analyst "thinks like an experienced data analyst." It is based on OpenAI's o3-mini AI model and optimized for advanced research. The tool uses "a reasoning chain to iteratively solve problems, reflecting human analytical thinking."
The digital assistant can use the Python programming language to run complex queries on data.
Distribution of Researcher and Analyst will begin in April as part of the new Frontier program.
