Google released Gemini 2.5 Pro (I/ O preview), a model update for the forthcoming Google I/O Congress, which was issued in advance and focused on upgrading coding-related capabilities, particularly in the areas of front-end development, UI construction, function calls, video understanding of integration code generation.
Main updates:
-
** Significant improvement in code capability** Better support function calls, module re-engineering, logical judgement
-
Structured understanding of “senior developers”
** Frontend and UI development capability upgrade**
-
WebDev Arena, number one.
-
Auto-generated beauty and well-structured web components (including CSS, responsive layout)
-
To recognize design styles and keep them consistent
** Video understanding + code generation integration**
-
Video content Interactive application (based on YouTube content)
-
Multimodular AI Capability Enhancement (VideoMME 84.8%)
** Designed to apply faster landing speed**
-
Available App for quick-generated UI+ action from simple hints
-
Example: Dietation Starter App
Experience optimization
-
Reduction in error rate, increase in function trigger rate
-
Compatibility with the old version interface, constant price, seamless upgrade
Detailed upgrades
1. ** Code upgrade**
-
Greater code generation and re-engineering: In an evaluation involving complex backend re-engineering tasks,** the abstract and structural selection capabilities of the “senior developer” level were demonstrated for the first time successfully through the Senior Developer Assessment for Cognition**;
-
For complex business logic adjustments, modular design, API rewriting, etc.
More reliable function call capability (Function Calling):
-
** Error rate drops, trigger rate increases** and developers can more easily construct a “Agentic Workflow” based on a combination of multi-step functions;
-
Developers of Gemini API access do not need to migrate and the system is automatically switched to a new version.
##2. Front-end development support for overall enhancement**
Gemini 2.5 Pro has leapfroged to **WebDev Arena’s number one ** in the front-end development field (the ranking for measuring human preferences).
-
Conversion of the design to the actual web component (HTML + CSS + JS);
-
Rapid generation of UI components and automatic adaptation to existing styles;
-
Understanding the visual elements of button interactions, fonts, margins, animations, etc., and restoring them to high-quality web pages;
-
Imitation of existing applications to achieve new functions (e.g. embedding video players in existing web pages).
Example: In Gemini 95 Starter App, the developers simply describe “ want to add a uniform video player”, and Gemini can read the existing design style automatically and generate usable components.
##3. Video+code: further integration of multi-model capabilities**
-
A high score on the VideoMME benchmark of 84.8 per cent, which is one of the leading “large models of video understanding” at present;
-
An interactive learning application may be generated automatically on the basis of a YouTube video**, including content extraction, interface construction and logical generation;
-
On behalf of AI, it has evolved from a language-driven code to a multi-model phase of a video-driven product.
Example: Video to Learing App of Google AI Studio demonstrates this new capability.
##4. Development process acceleration from conception to product** With Gemini 2.5 Pro, developers can quickly realize a product idea as an App with complete functionality and visual quality. Example product: “Dictation Starter App”:
-
Auto-generated** microphone animations, responsive layouts, button suspension effects**;
-
A high degree of professionalism in visual performance, not only in meeting functional needs;
-
Demonstrate the basic aesthetic perception and product awareness of the model.
Now you can use Gemini API in Google AI Studio to select Gemini 2.5 Pro (05-06) to build with Vertex AI.
-
The current Gemini 2.5 Pro preview is open for use at the same price as the old version (03-25)**;
-
The system has implicitly pointed the interface to the new version (05-06) and the developers do not need to migrate;
-
Model cards have been updated in a synchronized manner to record the details of the upgrades.