Gemini at Google I/0 2025: Lots of promises, but it will be necessary to be patient
Like every year at the same time, Google held her high mass, the Google I/O. This 2025 edition started on wheel hats, with a spectacular video introduction, entirely generated by Veo, the artificial video creation intelligence of the firm. A small demonstration of force in good and due form, before Sundar Pichai, the CEO of Google, goes on stage to open the ball of ads, very largely dominated by Gemini. But don’t ignite too much.
From the outset, the leader engaged in a self-brunt session, highlighting the performances of Gemini 2.5 Pro in the face of competition in the various benchmarks. He did not fail, in the form of a joke, to recall that Gemini had “finished Pokémon”, thus approaching, not of AG (general artificial intelligence), but of the “API” (artificial Pokémon Intelligence).
Beyond the anecdote, Sundar Pichai shared a striking figure: the use of Google models has been multiplied by almost fifty in the space of a year. He also clarified that if Gemini 2.5 Pro is currently the cutting edge for learning, Gemini Flash is the most efficient in terms of speed and cost.
Gemini 2.5 flash and improved audio capacities
The first concrete announcement concerned the new version of Gemini 2.5 Flash, presented as “better in almost all areas“That the previous iteration. This will be available during the summer. The most impatient can already test it within Google AI Studio, Vetex AI or via the Gemini application, probably for paid subscribers, although this point has not been explicitly mentioned.
Google then focused on improving audio interactions. Thus, Gemini 2.5 Pro and Flash now benefit from a native audio output, with the management of two distinct voices. This development promises more nuanced and expressive dialogues, the model being able to whisper or move from one language to another, as from Hindi to English, while retaining the same vocal intonation. This native audio output is accessible today via the Gemini API.
In the same vein, Gemini Live can now better distinguish the voice from the interlocutor from the background noises, allowing him to adapt his responses more relevantly. AI will even allow Translate conversations in quasi-time-real directly in a Google Meet videoconference.
AI is also more transversal, especially via further integration in Gmail.
Demos, still demos
We were then entitled to an impressive demonstration to solve mathematical problems by Gemini Diffusion. Another important announcement: the arrival of a new reasoning mode called Deep Think for Gemini 2.5 Pro. This is designed for very complex tasks. However, it will only be accessible at first only through the API for a limited number of “trust” developers, Google wishing to assess its capacities and safety.
The asynchronous coding agent, Jules, is now in public beta, accessible to all on registration.
World Models and the Astra project
Now let’s go to something sexier: the concept of “World Models”. The latter was illustrated by a conceptual video of the ASTRA project. We saw Gemini Live attending a user in the repair of a bicycle, finding a PDF manual, an explanatory video on Youtube, excavating in emails exchanged with the bike, and even going so far as to call the shop to order parts. A multitasking vision to say the least. And terribly frustrating. Because it is obviously a conceptive video, and nothing guarantees realization in the coming months. In addition, if the integration of Astra into a smartphone seems obvious, its potential would take on a completely different dimension in an accessory like connected glasses. Google has also touched this track by mentioning its partnership with AIRA.
Finally, Google has also unveiled a new “Google AI Ultra” package. And as much to tell you that he will not concern Grand-World, since it is offered at € 249.99 per month! For this price, you have access to the top of the top artificial intelligence in the brand (still happy). This does not only concern new gemini models, but also those for the generation of images (Imagen) and videos (VEO), as well as the least restrictive limitations possible in all services. And as a bonus, Google adds 30 TB of cloud storage in Google Drive and a Premium YouTube subscription. This new subscription is positioned in front of those of Openai and Anthropic, who also offer formulas so expensive for the most demanding … and wealthy users.
A feeling of unfinished
In the end, this keynote leaves a shared feeling on the Gemini front. If the announcements around Gemini are numerous and promising, most of them remain vague or reserved primarily for developers. Not to mention the new overpriced package. The general public will therefore have to be patient before being able to truly judge on pieces.
Want to save even more? Discover Our promo codes Selected for you.
