Join our daily and weekly newspapers for exclusive content on the latest updates and industry-composure AI coverage. learn more
Google A “universal AI assistant” is getting closer to its goal that can understand reference, plan and action.
today at Google I/OTech veteran announced a promotion for his Gemini 2.5 flash – it is now better in almost every dimension, including benchmarks for arguments, codes and long references – and 2.5 Pro, including an experimental enhanced argument mode, ‘Deep Think’, which Pro allowed to consider several hypotheses before reacting.
“This is our ultimate goal for the Mithun app: an AI that is personal, active and powerful,” said in a press pre-arid, CEO of Google Deepmind.
‘Deep Think’ score impressively on top benchmark
Google announced the Gemini 2.5 Pro-with a million-token reference window, considering its most intelligent model so far- in MarchAnd earlier this month (released “I/O” coding version with Hasabis Call it “The best coding model we have ever made!”).
“We have really been impressed as to what people have created, from converting the sketch into interactive apps to simulation of the entire cities,” Hasabis said.
He said that, based on Google’s experience with Alphago, AI model reactions improve when they are given More time to think. This inspired deepmind scientists to develop deep think, which uses the latest state -of -the -art research of Google in thinking and logic, including parallel techniques.
Deep Think has shown impressive score on the most difficult mathematics and coding benchmark including 2025 USA Mathematical Olympiad (State USA USAMOIt also moves forward LivcodbenchA hard benchmark for competition-tie-tier coding, and score at 84.0% MammuWhich tests multimodal understanding and logic.
“We are taking a little extra time to evaluate more frontier security and get input from safety experts,” Hasabis said. (Meaning: For now, it is available to reliable examiners through API for capacity, before the capacity is widely available.)
Overall, the new 2.5 Pro leads the popular coding leaderboard Webdave arenaWith an ELO score-which measures the skill level relative to players in two-Khiladi games such as chess -1420 (efficient from intermediate). It also leads to all categories Lamrena Leaderboard, which evaluates AI on the basis of human preference.
Since its launch, “We have actually been impressed (by users), from converting the sketch into interactive apps to simulation of the entire cities,” Hasbis said.
Gemini 2.5 Pro, Important Updates for Flash
Even today, Google announced a increased 2.5 flash, designed for its workharse models for speed, efficiency and low cost. 2.5 Flash has been improved in the board in benchmarks for Reasoning, Multimodality, Code and Long Reference – Hasabis said that it is “only the second” 2.5 Pro on the LMARENA leaderboard. The model is also more efficient using 20 to 30% less tokens.
Google is making final adjustments to 2.5 flash based on developer response; It is now available for preview in Google AI Studio, Wartax AI and Mithun apps. This will usually be available for production in early June.
Google is bringing additional capabilities for both Gemini 2.5 Pro and 2.5 flash, including indigenous audio output to create more natural conjunctive experiences, many speakers, thought summations and text-to-teach-speakes to support the budget.
With native audio input (in preview), users can run Gemini’s tone, pronunciation, and speaking style (think: directing the model melodramatic or model while telling a story). Like the project mergeer, the model is also equipped with the use of the tool, allowing it to discover from users.
Other experimental early voice features include emotional dialogue, which provides the model the ability to detect emotion in the user voice and respond properly; Active audio that allows it to tune the background conversations; Thinking in live API to support more complex tasks.
Many new-speaker features in both Pro and Flash support more than 24 languages, and models can quickly switch from one dialect to another. CTO of Google Deepmind, and Senior Director of Product Management at Google Deepmind, Tulsee Doshi said, “Text-to-speech is expressive and can catch microscopic nuances by whispering, Blog posted today,
In addition, 2.5 Pro and Flash now include Mithun API and Vartex AI idea summary. They “take the raw ideas of the model and arrange them in a clear format with information about header, key details, and model functions, such as they use the equipment,” Kavukuoglu and Doshi explain. The goal is to provide a more structured, streamlined format for the model’s thinking process and give users interaction with Gemini that are simple to understand and debug.
Like 2.5 flash, the Pro is also now equipped with the ‘Thinking Budget’, which gives the developers the ability to control the number of tokens using a model to think before reacting, or, if they prefer, turn off its thinking abilities completely. This capacity will generally be available in the coming weeks.
Finally, Google has added indigenous SDK support to the model reference protocol (MCP) definitions in Gemini API so that models can integrate more easily with open-source tools.
As Hasbis said: “We are living through a remarkable moment in history where AI is making a wonderful new future possible. This is tireless progress.”
