The world of generative artificial intelligence is burning and Google, one of the most prominent actors on the scene, continues to step on the accelerator with Gemini. In just a few days the company has chained a battery of updates and improvements that consolidate its Bet to bring AI everywhere -A all their parts, rather- with more power and more context than ever.
Among all those novelties, however, there is one that shines with its own light: Gemini 2.5 Prothe new model that Google defines without hesitation as the “smarter” they have created so far … which is not to say much; The strange would have not been to say it.
Gemini 2.5 Pro: Experimental, but with thought
What differentiates Gemini 2.5 from the previous version is not only the corresponding incremental improvement, but the default implementation and without noticing what Google calls and thinking modelan explicitly designed model to reason before generating an answer. It is not a novelty in the functioning of the LLM, but as results it offers more precision, better decisions and a capacity to understand that it goes beyond the usual.
The improvement, presumes Google, is noticeable in Benchmarks key as LMarena, where Gemini 2.5 Pro leads by far, as well as in demanding tests such as GPQA, Aime 2025 or Humanity’s Last Exam, where the model stands out without the need for additional techniques such as the majority voting. This not only speaks of a better basic architecture, but also of a post-training More refined, where Google has put special focus on reasoning and coding capabilities.
One of the most notable jumps is found in the field of development, where Gemini 2.5 pro exceeds version 2.0 in programming tasks, with an outstanding score of 63.8% in Swe-Bench Verified (the standard to evaluate the capable of acting as programming agents). It is not just writing code, but about transforming it, purifying it and even building visual applications from scratch with a simple prompt.
Finally, another of the great advances is in its context window, which reaches one million tokens and promises to double soon. This capacity allows Gemini to process massive information – from texts and videos to code repositories – and solve complex problems that require multiple sources and connections. In this regard, Gemini has no competition.
At the moment, but, Gemini 2.5 pro is only Available for Gemini Advanced usersthe payment plan, whether it is used through Google AI Studio or the Gemini application. Soon you will also come to seex AI and Google already prepare a pricing strategy that will scaling its use in production.
Gemini even in the soup … from Google
While the launch of Gemini 2.5 is the highlight that Google has left us this week in the field of AI, the company has not stopped deploying small – and not so small – updates that show that its road map for Gemini is to put it where it can have some utility, or where it is believed that it can have it.
One of the most striking is the new Gemini Live capacity to “see” in real time on Android devices. Thanks to the integration of Project Astra, some users can already share the screen or use the front camera for Gemini to interpret the visual environment and respond accordingly. It is nothing more than the first step of a function that will give a lot to talk about.
The arrival of gems to free users. Until now exclusive for Gemini Advanced subscribers, these Gems allow you to configure specific responses, tones, tones and objectives, very similar to how the Chatgpt custom GPT operates, but, at least for the moment, without their effectiveness.
Another interesting novelty is that, for the first time, Gemini can be used on the web without the need for a Google account. Accessing from the incognito mode, any user can interact with the 2.0 Flash model, although functions such as file loading or history remain exclusive for registered users.
