Google DeepMind once again raises the bar in the world of artificial intelligence: on March 25, 2025, the company unveiled Gemini 2.5—its smartest model yet, capable of “thinking” before responding. This announcement was shared on Google’s official blog.
The new release, featuring Gemini 2.5 Pro, promises to become a true breakthrough in AI, demonstrating exceptional results in mathematics, programming, and complex analytical tasks.
Gemini 2.5 Pro has already set new standards: without resorting to costly methods such as “majority voting,” it leads benchmarks in mathematics and science, including GPQA and AIME 2025, and has scored 18.8% on Humanity’s Last Exam, a set of challenges designed by experts to test the limits of human knowledge. Especially impressive is the progress in programming: on the SWE-Bench Verified test, the model achieved a 63.8% success rate, creating complex web apps and even video games from simple textual requests. Imagine typing “make a game about space,” and Gemini 2.5 Pro delivers executable code!
The model has also become a true master of “agentic code,” allowing the creation of applications that operate autonomously and execute code-editing tasks with accuracy unmatched by previous versions. Compared to Gemini 2.0, this latest iteration represents a significant leap forward in coding capability, and Google promises further improvements in the near future. As a bonus for developers, Gemini 2.5 is already available via Google AI Studio and Gemini API, and for Gemini Advanced users through a monthly subscription priced at $20.
