Google introduces an upgraded preview of Gemini 2.5 Pro, its most intelligent model yet, designed for enterprise-scale applications and soon to be generally available. The model shows significant performance improvements, including a 24-point Elo score increase on LMArena (leading at 1470) and a 35-point jump on WebDevArena (leading at 1443). It excels in coding benchmarks like Aider Polyglot and challenging tests such as GPQA and Humanity's Last Exam, which assess math, science, knowledge, and reasoning skills.
Google improved the model's style and structure based on user feedback, making responses more creative and better formatted. These enhancements address concerns about output quality and presentation.
Developers can access the upgraded preview through the Gemini API via Google AI Studio and Vertex AI, which now include thinking budgets to help manage cost and latency. The enhanced model is also rolling out in the Gemini app, expanding access beyond the developer community.