Google unveils a family of new generation AI reasoning models

Tuesday, Google unveiled Gemini 2.5, a new family of AI reasoning models that stops to “think” before answering a question.

To launch the new family of models, Google Lance Gemini 2.5 Pro Experimental, a multimodal and AI reasoning model which, according to the company, is its smartest model to date. This model will be available Tuesday in the company’s developer platform, Google AI Studio, as well as in the Gemini application for subscribers to the $ 20 IA plan per month of the company, Gemini Advanced.

In the future, Google says that all of its new AI models will have reasoning skills.

Since Openai launched the First IA reasoning model in September 2024O1, the technological industry has run to equal or overcome the capacities of this model with theirs. Today, Anthropic, Deepseek, Google and Xai all have models of AI reasoning, which use an additional calculation power and time to check the facts and reason with problems before providing an answer.

The reasoning techniques have helped AI models to reach new heights in mathematics and coding tasks. Many in the world of technology believe that reasoning models will be a key element of AI agents, autonomous systems that can perform tasks without human intervention. However, these models are also more expensive.

Google has already experienced AI’s reasoning models, previously published a “thought” version of Gemini in December. But Gemini 2.5 represents the most serious attempt of the company to date on the drop in the series of “O” models of Openai.

Google claims that Gemini 2.5 PRO surpasses its preceding border AI models, and some of the main models of competitors, on several landmarks. More specifically, Google says that it designed Gemini 2.5 to excel in creating web applications and visually convincing agent coding applications.

During an evaluation measuring code editing, called to help Polyglot, Google says that Gemini 2.5 PRO scores 68.6%, outperforming higher AI models from Deepseek Openai, Anthropic and Chinese AI LAB.

However, on another test to measure software capabilities, Swe-Bench checked, Gemini 2.5 Pro brand 63.8%, outperforming O3-Mini of Openai and R1 from Deepseek, but underperforming the Sonnet Claude 3.7 of Anthropic, which marked 70.3%.

During the last examination of humanity, a multimodal test made up of thousands of Crowdsourced questions relating to mathematics, human sciences and natural sciences, Google says that Gemini 2.5 pro scores 18.8%, operating better than most rival flagship models.

To start, Google says that Gemini 2.5 Pro is shipped with a 1 million token context window, which means that the AI ​​model can take about 750,000 words in a single time. It’s longer than the whole series of books “Lord of the Rings”. And soon, Gemini 2.5 Pro will support the double of the entry length (2 million tokens).

Google did not publish the API Prize for Gemini 2.5 Pro. Society says it will share more in the coming weeks.