Google introduces Gemini 2.5 Pro, its most capable AI yet

Google has introduced Gemini 2.5 Pro, an experimental artificial intelligence model that can analyze a problem before responding. The company describes it as its most advanced model yet and says it debuts in first place on LMArena, a ranking based on user preferences, with a significant lead.

The main development is not just its ability to generate text, images, or code. Gemini 2.5 incorporates reasoning directly into the model: it can break down a complex task, consider the context, and arrive at an answer step by step before showing you the result.

What it means for Gemini to reason

In practice, this should help with tasks where recognizing patterns is not enough. For example, it can analyze a math problem, compare information from multiple sources, or modify an application based on detailed instructions.

Google had already experimented with this idea in Gemini 2.0 Flash Thinking. With Gemini 2.5, the company combines a more powerful base model with improved post-training. Its goal is to include these capabilities in all its future models, including those that work as agents capable of completing tasks with less human intervention.

Google's published results place Gemini 2.5 Pro among the top performers in several math, science, and general knowledge tests. Highlights include:

• 18.8% on Humanity's Last Exam, an evaluation created by specialists to measure knowledge and reasoning on difficult questions.
• Leading results on tests such as GPQA and AIME 2025, which focus on science and mathematics.
• 63.8% on SWE-Bench Verified, a test that measures the ability to solve real-world programming problems with the help of a customized agent.

These figures come from specific evaluations. They do not guarantee that the model will always be correct. Performance can change depending on the task, the instructions, and the information it receives.

A model built for programming and large amounts of context

Gemini 2.5 Pro stands out especially in programming. Google says it can create visual web applications, transform existing code, and develop applications capable of taking actions on their own. The company also showed how the model generates the executable code for a video game from a single line of instructions.

The model includes a 1 million-token context window. In other words, it can retain and analyze a very large amount of information within the same conversation or task. Google is preparing an expansion to 2 million tokens.

It also supports text, audio, images, video, and complete code repositories. For you, this means it could analyze, for example, a project's documentation, its code files, and a recording of an error without forcing you to split everything into many separate queries.

Where you can use it

Gemini 2.5 Pro Experimental is already available in Google AI Studio and in the Gemini app for people with a Gemini Advanced subscription. It will also come to Vertex AI, Google's platform for businesses and developers, in the coming weeks.

Google has not yet announced pricing for intensive production use. The company says it will publish that information in the coming weeks, along with higher usage limits for customers who need to integrate it into their own services.

The important question now is whether its strong performance in tests carries over to everyday work: programming, research, summarizing long documents, or coordinating tasks across multiple tools. Gemini 2.5 signals Google's commitment to models that do not just respond quickly, but spend more effort solving complex problems before answering.