Google has announced the Gemini 2.5 Pro AI model, calling it “its smartest model yet.” As part of the Gemini 2.5 family, it surpasses previous versions in data analysis, programming, and complex problem-solving. A key highlight is its expanded context window, supporting up to 1 million tokens, with plans to extend it to 2 million soon.
Enhanced Reasoning and Performance
A defining feature of Gemini 2.5 Pro is its ability to reason before delivering answers. Unlike the Gemini 2.0 Flash Thinking model, this version no longer displays its thought process by default. However, users can manually activate the “think out loud” feature to see how the AI arrives at its conclusions. Thanks to an improved base model and post-training refinements, Gemini 2.5 Pro has achieved top rankings in multiple evaluations, including the LMArena rating, mathematics (AIME 2025), and science (GPQA diamond).
In a specialized test, Humanity’s Last Exam—designed to assess AI’s logical limits—the model set a record with an 18.8% score without additional tools. It also made notable progress in programming, particularly in web application development and code editing, adds NIXsolutions.
Expanding Capabilities and Availability
In software development, Gemini 2.5 Pro excelled in the SWE-Bench Verified test, achieving 63.8% with a special agent approach. The model’s multimodal abilities allow it to process text, audio, images, video, large datasets, and even full code repositories.
Currently, Gemini 2.5 Pro is accessible to paid subscribers and developers in test mode. It will be integrated into Vertex AI in the coming weeks, followed by the introduction of a pricing model for large-scale projects. We’ll keep you updated as more details emerge.