Google's Gemini 2.5 Pro Preview is the best on the web!

I/O Eve.googlesudden outburst Gemini 2.5 Pro preview version

Last night, Google announced the launch of the I/O conference ahead of the Gemini 2.5 Pro Preview (I/O version). As its flagship Gemini An upgraded version of the 2.5 Pro AI model, which the search giant claims to have topped a number of mainstream benchmarks.

The Gemini 2.5 Pro Preview (I/O version) is available through the Gemini API and Google's Vertex AI and AI Studio platforms for the same price as the pre-upgrade Gemini 2.5 Pro model. The preview has also been introduced into Google's Gemini chatbot app, which is available primarily for web and mobile devices.

google DeepMind (scientific research organization) CEO Demis Hassabis 在 X 上发帖称：“非常高兴能与大家分享我们迄今为止构建的最佳编程模型！今天，我们发布了 Gemini 2.5 Pro Preview 的‘I/O 版’，其编程能力得到了大幅提升。This version was ranked #1 in the LMArena Programming category and also topped the WebDev Arena charts!."

The release of the models coincides with Google's annual I/O developer conference (hence the name "I/O Edition"). Google is expected to release a series of models and AI-driven tools and platforms at the conference. In this brutal AI race, Google is struggling to capture market share and user attention; competitors such as OpenAI and xAI are also about to release powerful models of the same kind.

What can the latest version of the model actually do?

According to Google, the Gemini 2.5 Pro Preview (I/O release) delivers "significant" improvements in coding and building interactive web applications. In addition, the new version of the model also excels at tasks such as code transformations (i.e., modifying a piece of code to achieve a specific goal) and code editing.

In a blog post, Google noted that Gemini 2.5 Pro Preview (I/O version) topped the Web Arena leaderboard, which measures a model's ability to create beautiful and powerful Web applications, by 147 Elo points over the previous version. The chart measures human preference for a model's ability to build beautiful and powerful Web applications. It also continues to build on its strong foundation in native multimodality and long context.

WebDev Arena, developed by LMArena, is a real-time AI coding competition where models go head-to-head in web development challenges

Video to Code

It's worth noting that the new version of the model is also a top performer in video comprehension, scoring 84.8% on the popular benchmark VideoMME.

Combined with the encoding capabilities, Gemini 2.5 Pro Preview enables a whole new process that was not possible with previous versions. With greater video comprehension and a refined user interface, the upgraded Gemini 2.5 Pro model offers a more feature-rich experience than the simple previous example.

In a blog post, Google wrote, "For developers working with Gemini 2.5 Pro Preview, this new release will not only improve programming performance, but will also address a number of key pieces of feedback from developers, including reducing errors in function calls and increasing the trigger rate of function calls. By default, the model remains both well maneuverable and truly reflects the aesthetic pursuit of Web development."

Make feature development easier

Gemini 2.5 Pro Preview excels at front-end web development. Implementing the new functionality means that users need to manually drill down into the design files, check the components to match style attributes such as colors, fonts, inner margins, outer margins, and borders, and then manually write the required CSS code to accurately replicate these visual attributes. Imagine using Gemini 2.5 Pro Preview in the IDE and having the model generate new functionality, such as adding a video player that is styled the same way as the rest of the applications in the Gemini 95 Starter App.

Make it easier to get ideas off the ground

Gemini 2.5 Pro Preview makes it easy and efficient to get ideas off the ground, offering both powerful utility features and an elegant user interface. Take, for example, the newly upgraded Dictation Starter app, which is built on the latest models and demonstrates the benefits of Gemini 2.5 Pro Preview.

What's noteworthy is the attention to detail: dynamic wavelength animations, smooth responsive layouts, and clever button hover interactions. The model is a natural fit for modern Web development, ensuring visual aesthetics while focusing on ease of use, helping developers quickly turn concepts into real-world Web applications.

For example, Gemini 2.5 Pro Preview intelligently designs microphone UI animations for dictation applications and automatically generates the corresponding code, significantly improving development efficiency.

What do outsiders say?

The release of Gemini 2.5 Pro Preview has generated a lot of buzz in the community, and Michael Truell, CEO of AI coding tool Cursor, couldn't help but praise Gemini 2.5 Pro Preview for taking coding even further.

"We are very excited about the latest Gemini 2.5 Pro Preview, which builds on its strong real-world programming capabilities. We have observed internally a significant reduction in the number of failed calls to the tool with the new model, and we are confident that users will find that this improvement makes 2.5 Pro Preview more efficient in Cursor than ever before."

On Hacker News, some users have argued that the Gemini 2.5 series is more reliable than other models, but still can't compete with human developers:

"Formerly, when programming with other models, I often ran into a headache: they sometimes generated APIs that simply didn't exist. in contrast, the Gemini 2.5 series (both the Pro and Flash versions) performs much better in this regard, and more reliably than any other model I've tried.

Of course, it still has some obvious limitations. No matter how much the cue words are optimized, the current model is still no match for human developers in terms of abstract thinking and system architecture. But even so, I've found that Gemini has been able to replace many of my daily searches and StackOverflow queries, significantly improving my programming productivity."

A user echoed these sentiments, arguing that even though the Gemini 2.5 Pro Preview is good enough for intelligent coding, it is still not comparable to humans at this point in time, however, he believes that it is only a matter of time before AI surpasses human beings in terms of coding capabilities.

"It's clear that it's only a matter of time before humans are overtaken in the realm of codesign (whether this is 1 year or 5 years from now doesn't really matter). Instead of dwelling on these trends that can't be changed, we should focus on more valuable questions: what can we do in this new world that's coming? We need more constructive ideas, and this is the perfect place to facilitate those discussions."

A developer who uses large models daily to aid programming said:

"I use the Large Language Model (LLM) every day to aid programming. The programming power of LLMs has certainly improved significantly over the years, but it has focused primarily on the ability to map 'natural language to code'. While this is a powerful capability, it should be used with care: the context needs to be carefully managed to keep the model focused, and the model must be proactively guided to take into account key factors such as performance optimization and system architecture. I continue to have reservations about the reasoning capabilities of large models. This is not to dismiss their value, but to recognize their inherent limitations. I believe that to achieve true human-like intelligence, we may need to explore a completely different technology path than LLM."

However, there are X users who believe that Gemini 2.5 Pro Preview has only slightly improved coding skills. Everything else has decreased slightly. Therefore, it's just a big model made for coding and not suitable for general use (compared to previous versions).

Reference Links

https://blog.google/products/gemini/gemini-2-5-pro-updates/

This article is from WeChat"AI Frontline.", by Winterberry, Nuclear Cola