Google announces Gemini 2.5, its most-intelligent AI language model: What it offers over the older Gemini 2.0

Gemini 2.5

Google has unveiled Gemini 2.5, an enhanced AI language model with improved reasoning, coding, and multimodal capabilities. The new model, introduced as Gemini 2.5 Pro Experimental, surpasses previous versions, demonstrating state-of-the-art performance across various AI benchmarks.

One of the primary advancements in Gemini 2.5 is its improved reasoning ability. Unlike its predecessor, the model can analyze information more efficiently, incorporate context, and draw logical conclusions. Google attributes this improvement to a combination of an enhanced base model and refined post-training techniques. The model leads benchmarks in mathematics and science, scoring 18.8% on Humanity’s Last Exam, a dataset designed to assess AI’s ability to handle complex knowledge-based questions.

Gemini 2.5 marks a significant leap in coding capabilities compared to Gemini 2.0. The model excels in creating web applications, agile code applications, and code transformation tasks. It achieves a 63.8% score on SWE-Bench Verified, an industry standard for evaluating coding agents. Google also showcased an example where the model produced executable code for a video game from a single-line prompt.

Building on the multimodal capabilities of previous Gemini models, Gemini 2.5 features an expanded context window of 1 million tokens, with an upcoming upgrade to 2 million tokens. This enhancement enables the model to process and comprehend large datasets, images, videos, and code repositories more effectively. Such improvements aim to assist developers and enterprises in handling complex problem-solving tasks across multiple data sources.

Gemini 2.5 Pro is now available in Google AI Studio and the Gemini app for Gemini Advanced users. It will be accessible via Vertex AI in the coming weeks, with additional pricing plans set to be introduced. Google plans to integrate advanced reasoning capabilities into all future Gemini models, ensuring they remain competitive in AI-driven applications.