Google launches new Gemini 2.5 Pro model featuring improved capabilities

Today, Google unveiled the enhanced Gemini 2.5 model, showcasing notable advancements in performance across several prominent AI benchmarks. The tech giant specifically emphasized that the new Gemini 2.5 Preview 06-05 “Thinking”model excels in key areas such as coding, mathematics, science, and logical reasoning.

Our latest Gemini 2.5 Pro update is now in preview. It’s better at coding, reasoning, science + math, shows improved performance across key benchmarks (AIDER Polyglot, GPQA, HLE to name a few), and leads @lmarena_ai with a 24pt Elo score jump since the previous version. We also… pic.twitter.com/SVjdQ2k1tJ

— Sundar Pichai (@sundarpichai) June 5, 2025

The recent announcement builds upon the developments introduced last month during Google I/O, where an initial upgraded version of the Gemini 2.5 Pro model was released. This latest iteration not only enhances benchmark scores but also addresses user feedback, which has led to improved stylistic and structural elements in responses, resulting in more creative and well-formatted outputs for users.

Below, you’ll find a comparative analysis of benchmarks against other leading AI models.

The accompanying table highlights that the Gemini 2.5 Pro preview model now stands as State-of-the-Art (SOTA) in coding benchmarks such as AIDER Polyglot. Furthermore, it achieved SOTA status in the GPQA and Humanity’s Last Exam (HLE) benchmarks, which evaluate capabilities in mathematics, science, general knowledge, and reasoning skills. Remarkably, during real-world evaluations, this latest 2.5 Pro model exhibited a 24-point improvement on LMArena, and a significant 35-point surge on WebDevArena, reaching a score of 1443.

Developers interested in utilizing this advanced Gemini 2.5 Pro preview model can access it through the Gemini API available at Google AI Studio and Vertex AI. In addition, general users will be able to access this model via the Gemini app. Google has confirmed that the Gemini 2.5 Pro model will be made widely available in a few weeks, paving the way for integration into enterprise-scale applications during production.

Source & Images