Join our daily and weekly newsletters to obtain the latest updates and exclusive content on the coverage of the industry leader. Get more information
Google is approaching its objective of a “universal AI assistant” who can understand the context, plan and take action.
Today, in Google I/O, the technological giant announced improvements to its Gemini 2.5 Flash, now it is better in almost dimensions, including reference points for reasoning, code and long context, and 2.5 PRO, including a fashion with alleviated multiple hypotheses before responding.
“This is our ultimate goal for the Gemini application: an AI that is personal, proactive and powerful,” said Demis Hassabis, CEO or Google Deepmind, in a previous press letter.
‘Deep Think’ gets impressive at the best reference points
Google announced Gemini 2.5 that considers its smartest model so far, with a million context token in March, and launched its “E/S” coding edition at the beginning of this month (with the so-called Hassabis! “).
“We have a really impressed leg for what people have created, from converting sketches into interactive applications to simulating whole cities,” said Hassabis.
He pointed out that, according to Google’s experience with Alphago, the answers of the AI model improve when they are given More time to think. This led Deepmind scientists to develop Deep Think, who uses Google’s latest research and reasoning, including parallel techniques.
Deep Think has shown impressive scores at the most difficult mathematics and coding points, including the United States mathematical Olympiad in 2025 (Usamo). It also leads to LiveCodeBench, a difficult reference point for the coding at the level of competition, and obtains 84.0% in MMMU, which proves multimodal understanding and reasoning.
Hassabis added: “We are taking a bit or extra time to perform more border security evaluations and obtain more information from security experts.” (Meaning: As for now, it is available for confidence testers through the API to obtain feedback before the capacity is widely aviable).
In general, the new 2.5 Pro leads the popular webdev Arena coding classification table, with an Elo score, which measures the relative skill level of players in games of two players such as the chess of 1420 (intermediate to competent). It also leads in all categories of the LMarena classification table, which evaluates AI based on human preference.
Since its launch, “we have the leg really impressed by what [users have] Created, from converting sketches into interactive applications to simulating whole cities, “said Hassabis.
Important Gemini 2.5 Pro updates, flash
Also today, Google announced an improved flash 2.5, considered its battle horse model designed for speed, efficiency and low cost. 2.5 Flash has improved in all areas in reference points for reasoning, multimodality, code and long context. Hassabis said it is “only” to 2.5 Pro in the LMarena classification table. The model is also more efficient, using 20 to 30% less tokens.
Google is making final adjustments to 2.5 flash based on developer feedback; It is now available for previous view on Google Ai Studio, VerTex AI and in the Gemini application. It will generally be available for production in early June.
Google is providing additional capabilities to Gemini 2.5 Pro and 2.5 Flash, including native audio production to create more natural conversation experiences, voice text to admit multiple speakers, thought summaries and thought budgets.
With the native audio entry (in preview), users can direct the tone, accent and the style of talking about Gemini (think: direct the model as melodramatic or Maudlin when it tells a story). Like Project Mariner, the model is also equipped with the use of tools, which allows you to search the being of users.
Other experimental characteristics of early voice include affective dialogue, which gives the model the ability to detect emotion in the user’s voice and respond properly; Proactive audio that allows you to adjust the background conversations; And think about live API to support more complex tasks.
The new characteristics of multiple speakers in PRO and Flash admit more than 24 languages, and models can quickly change from one dialect to another. “The text to the voice is expressive and can capture subtle nuances, such as whispers,” wrote Koray Kavukcogu, CTO or Google Deepmind, and Tulsee Doshi, senior director of Product Management on Google Deepmind, in a blog published today.
In addition, 2.5 Pro and Flash now include thought summaries in the Gemini API and Vertex AI. These “take the gross thoughts of the model and organize them in a clear format with headers, key details and information on model actions, such as when they use tools,” Kavukcogl and Doshi explain. The objective is to provide a more structured and simplified format for the model thinking process and provide users interactions with Gemini who are simpler to understand and purify.
Like 2.5 Flash, Pro is also equipped with ‘Thought Budgets’, which gives developers the ability to control the number or think about thinking before responding, or, if they prefer, turn off the abilities completely. This capacity will generally be available next week.
Finally, Google has added the native SDK support for the definitions of the model context protocol (MCP) in Gemini’s API so that the models can be more easily integrated with open source tools.
As Hassabis said: “We are living a remarkable moment in history where AI is making possible an incredible future. It is a bone rolling.”
