- Google Gemini 2.5 Pro gets a new Deep Think model
- Deep Think allows Gemini to consider several reasoning paths before responding
- Deep Think will improve the accuracy of Gemini on Mathematics and Complex Code
Google adds an additional brain to Gemini with a new deep thinking mode. The company has unveiled the latest option for Google Gemini 2.5 PRO to Google I / O 2025, showing exactly what its AI can do with additional depth.
Deep Think essentially increases “the spirit” of Gemini’s AI with additional brains. Gemini in in -depth thinking do not simply spit a response to a request as quickly as possible. Instead, he performs several possible lines of reasoning in parallel before deciding how to react. It is like the equivalent of the AI to look in both directions, or to reread the instructions before building a piece of furniture.
And if Google’s tests are to be passed, Deep Think’s Brainpower works. It occurs at a high level level of the Olympiad of American mathematics 2025, leaving at the top of the competitive programming test Livecodebench, marking 84% incredibly high on the MMMU popular, a kind of decathlon of multimodal reasoning tasks. Deep Think is not yet widely available. Google unrolls it to trusted testers only for the moment. But, probably, once all the folds are pushed, everyone will have access to the deepest thoughts of Gemini.
Gemini shines on
Deep Think fits perfectly into the rest of the growing range of Gemini 2.5 and the new features arriving for its different models in the API used by developers to integrate gemini into their software.
For example, Gemini 2.5 Pro now supports the generation of native audio. This means that he can speak to you. The discourse has an “emotional dialogue” functionality, which detects emotional changes in your tone and adjusts accordingly. If you seem stressed, Gemini could stop talking as a patient customer service agent and respond more as an empathetic and thoughtful friend (or at least how the AI interprets such an answer). And it will be better to know when to speak at all thanks to the new proactive audio functionality, which filters background noise, so Gemini only sounds when he is sure to speak to him.
Associated with new security guarantees and the next features of the Mariner Computer-Us project, Gemini 2.5 is very strongly tried to be AI in which you trust not only with your calendar or code, but with your book narration or an entire operating system.
Another element developing on Gemini 2.5 is what Google calls a “reflection budget”. Previously unique to Gemini 2.5 Flash, the reflection budget allows developers to decide how much the model should think before responding. It’s a good way to make sure you get a full response without spending too much. Otherwise, Deep Think could give you just a taste of his reasoning, or give you everything and make it too expensive for any follow-up.
In the event that it is not clear what these thoughts imply, Gemini 2.5 Pro and Flash will offer “summaries of reflection” for the developers, a document showing the exact details of what the AI was doing in terms of information application via its reasoning process, so that you can really look inside the brain of AI.
All this signals a pivot of models which speak simply quickly to underline those who can reason more deeply, although slower. Deep Think is part of this change towards deliberate and lying reasoning. It is no longer just to try to predict the following word, it applies this logic to ideas and the process even to find answers to your questions. Google seems eager not only to make gemini not only capable of recovering answers, but to understand the form of the question itself.
Of course, AI reasoning still exists in a space where a perfectly logical response could be delivered with a random side of nonsense, regardless of the impressive reference scores. But you can start to see the shape of what is coming, where the promise of a real “co -pilot” materializes.
