Inside Google’s AI leap: Gemini 2.5 thinks deeper, speaks smarter and codes faster

Google The context is approaching a “universal AI assistant” that can understand, plan and move.

Today Google i / OThe Tech giant announced those who entered the twins 2.5 – an experienced mode of thought, “Deep Thinking” and 2.5 Pro, “Deep Think”, to calculate numerous assumptions before answering.

“This is our last goal for the Gemini App: AI, Google Deeprabis, CEO, which is a personal, active and powerful, said in a short press in advance.

‘Think deep’ in the best criteria effectively

Google Gemini 2.5 Pro – Still announced the smartest model with a million-token context window – In Marchand released “I / o” coding publication At the beginning of this month (with Hassabis) call “The best coding model we’ve built!”).

“We were really affected after people create interactive applications to imitate all cities to imitate all cities,” Hassabis said.

He noted that based on Google experience AlbligoAI is improving when the model answers are given more time to think. This uses the latest advanced research, including Google’s latest advanced investigation, including parallel methods, to develop Deepmind scientists.

Deep thinking showed impressive points in the most difficult mathematics and coding criteria, including 2025 US Math Olympiad (-Inking). Also leads LivecodebenchA difficult benchmark for competitive level coding and 84.0% scores MacwannessMultimodal understanding and testing the thought.

Hassabis added: “We allocate a little extra time for more border security assessments and we are taking a little time to get more access to security professionals.” (Meaning: So far, it is possible for those who have reliable tests through the API before the expansion of the opportunity.)

In general, the new 2.5 Pro is headed by a popular coding leader board Webdev arenaWith an Elo account – the level of relative skill levels of players – 1420 (experienced intermediate) in two player games like Chess. It is also a leader among all categories Lmana Leader board assessing AI based on man’s choice.

Since the day it was started, we were really affected by what [users have] To imitate all cities, the sketches were created to convert interactive applications, “Hassabis said.

Gemini 2.5 Pro, Important News for Flash

In addition, it announced an advanced 2.5 flange, taking into account the model of Google, speed, efficiency and workplaces designed for a low price. 2,5 Flash, justification, multimolism, code and long context – Hassabisin Library Leaders’ Leaders’ Leaders’ board was “only second” on the board. The model is more efficient using 20 to 30% less verses.

Google makes the latest adjustments for 2,5 flash based on the developer’s review; Now Google AI Studio is available to advance in the application of Vertex AI and Gemini. This will generally be available for production in early June.

Google brings additional opportunities, including native sound speeches, including more natural scratch experiments, including more natural scratch practices, many speakers, thinking and meditation budgets, including 2.5 flash.

With native audio access (preview), users can hold twins, emphasis, emphasis and speaking steering (think: Model the model for being melodramatic or maudlin while talking. Approve Project MarinerThe model is also equipped with tool usage that allows users to look for on behalf of users.

Another experimental early sound features include an affective dialogue that allows the model to detect emotions in the user’s voice and answer properly; Active voice that allows you to adjust background conversations; And think to support more complex tasks in the live API.

The features of new numerous spinfacies in both Pro and Flash support can be rapidly from 24 languages ​​and models to another. “Text-to-speech expression and delicate nuances such as whispering,” Koray Kavukcuoglu, Google Deepmind “and Tulsee Doshi, Tulsee Doshi, Tulsee Doshi, Tulsee Doshi, Tulsee Doshi, Tulsee Google Deprmind wrote in ADO Blog was sent today.

In addition, 2.5 Pro and Flash now include the Gemini API and thoughtful summaries in Vertex AI. These are “Kavukcuoğlu and Doshi”, “Kavukcuoğlu and Doshi”, “Kavukcuoğlu and Doshi” explain the model.

2.5 Like Flash, the Pro is equipped with ‘thinking budgets’, which allows you to manage the number of a model that responds to developers, or completely turn off the thinking capacity. This ability will be available in the future weeks.

Finally, Google, Models API added native SDK support for model context protocol (MCP) recipes that models can be more easily integrated with open source tools.

As Hassabis puts, “We live through a remarkable progress in the history of the AI’s incredible new future. This was a ruthless progress.”