Before answering, they pause to reflect.
Introducing Gemini 2.5, our most intelligent AI model.
The first version, an experimental 2.5 Pro release, delivers the best performance possible in mathematics and science.
Google (@Google), March 25, 2025
Gemini 2.5 Pro is an experimental version. The most intelligent multimodal reasoning network to date is the “multimodal reasoning neural net.” The Gemini app is $20 per month for Advanced subscribers and available in both the Google AI Studio platform for developers as well as the Google AI Studio for the developer platform.
In a recent blog, the company said that “in Gemini 2.5 we achieved a higher level of performance through combining an improved base model and advanced post-learning.”
In the future, the company said that all its models would be able to reason.
Gemini 2.5 Pro beats OpenAI, Anthropic, and DeepSeek in the Aider polyglot test that measures programming abilities.
Google lost the benchmark SWE Bench Verified, which tests the software developer’s abilities, to Claude 3. 7 Sonnet by 63.8% against 70.3%.
Gemini 2.5 pro scored 18,8% in the Humanity’s Last Exam multimodal test, consisting of 1,000 questions on math, science and humanities. This is higher than the scores of the competitors.
Gemini 2.5 Pro comes with a 1,000,000 token contextual window. The model can ingest approximately 750,000 sentences at once. The plan is to increase the number in the near future.
For “reasoning” AI to be accurate, it requires more computational power and time. OpenAI was the first to show a “thinking neural network” – it came out with o1 on September 20, 2024. Others have been busy developing their own AI solutions. DeepSeek is one of them, as are Anthropic and Alibaba.
In March of this year, Google released a range of AI features and products for the Healthcare sector.
Gemini, the company’s AI assistant, was previously programmed to take into consideration a user’s past search histories and provide more accurate answers.
Microsoft releases two AI agents to ‘deep-research’
Microsoft will be adding two AI “reasoning agents” to Microsoft 365 Copilot that can analyze huge amounts of data and offer highly qualified expertise – Researcher and Analyst.
The multi-step process of analysis is made easier with Researcher. This tool combines OpenAI Deep Research Mode for deep search with “advanced searching capabilities” in Microsoft 365 Copilot.
[embedded content]
The company says that users can create a go-to market strategy with Researcher. It will analyze competitor data and business information on the Internet to identify new opportunities, or produce a performance report every quarter.
Researchers can leverage data from AI tools, applications and agents like Confluence and ServiceNow.
The Analyst thinks “like an experienced data analyst.” The model is OpenAI’s o3 mini AI and has been optimized for advanced researchers. It uses a “reasoning chain” to solve problems iteratively, reflecting analytical thinking.
[embedded content]
Python is a programming language that can be used by the digital assistant to perform complex data queries.
As part of Frontier’s new program, the distribution of Researcher and Analyst will start in April.
