.png)
May 7, 2025
Shortly before the Google I/O 2025 developer conference, Google has released early access to an improved preview version of Gemini 2.5 Pro. Google justifies the early release with the "great enthusiasm" of the developer community and the "impressive results" already achieved with the previous version of Gemini 2.5 Pro.
The update goes beyond mere surface improvements. Google emphasizes the "overwhelmingly positive feedback" on the original version, particularly regarding its programming capabilities and the processing capacity of various data types. Developers can look forward to "significant improvements for frontend and UI development." The improvements also affect complex code transformations, detailed code editing, and the development of agent workflows – AI systems that can perform tasks largely autonomously.
The performance increase is reflected in the results: Gemini 2.5 Pro now leads the WebDev Arena rankings, surpassing its predecessor by a considerable 147 Elo points. This ranking evaluates the ability of AI models to create functional and appealing websites. The progress indicates a significant leap in the AI's ability to design user-friendly web experiences. Google highlights that this development already powers the "innovative code agents of Cursor" and promotes collaboration with companies like Cognition and Replit at the "frontiers of agent programming."
Executives from these companies confirm the power of the updated model. Michael Truell, CEO of Cursor, reports a significant reduction in errors in function calls. Michele Catasta, President of Replit, describes Gemini 2.5 Pro as the best Frontier model in the ratio of "capability to latency." Silas Alberti from the founding team of Cognition praises the "leading performance" and the model's ability to "make correct decisions and choose good abstractions."
In addition to its programming capabilities, Gemini 2.5 Pro scores with its ability to understand various media types and process large amounts of data. Google proudly points to the "top performance in video understanding" and a score of 84.8% in the VideoMME benchmark. The combination of video and programming capabilities opens up new possibilities. An example is the "Video to Learning App" in Google AI Studio, which generates an interactive learning app from a YouTube video.
For developers, this means significant relief in the development of new features, especially in the frontend area. Gemini 2.5 Pro can be used in code editors, for example, to generate new elements that seamlessly integrate into the design of existing apps. Google demonstrates this using a dictation app created with the updated model, which impresses with "wavelength animations, responsive design, and subtle button hover effects." The model has a distinct sense of aesthetic web development and simultaneously allows developers to quickly translate concepts into functioning web apps.
The updated Gemini 2.5 Pro is available via the Gemini API in Google AI Studio and for businesses via Vertex AI. It will also be integrated into the regular Gemini app and enable features like Canvas and "Vibe Code," which allows anyone to create interactive web apps by simply entering text. For existing users of Gemini 2.5 Pro, the upgrade is automatic and at no additional cost. Google has also taken specific developer feedback into account and, among other things, reduced the error rate in function calls.
By providing the improved features early, Google accelerates the feedback process and optimizes the development of its tools. With a focus on practical programming – from complex backend revisions to UI generation – Gemini 2.5 Pro is evolving into a valuable tool for software developers.
