During the Google I/O 2025 event on Tuesday, the company unveiled a range of innovative features for its Gemini 2.5 series of artificial intelligence models. The Mountain View technology leader introduced a sophisticated reasoning mode called Deep Think, powered by the Gemini 2.5 Pro model. Additionally, Google presented Native Audio Output, a new feature that enables more natural and human-like speech, which will be accessible through the Live application programming interface (API). Furthermore, enhancements such as thought summaries and thinking budgets are being incorporated into the latest Gemini models for developers.
Gemini 2.5 Pro Tops LMArena Leaderboard
A recent blog post from Google outlined the new features and improvements set to roll out for the Gemini 2.5 AI model series in the coming months. The company had earlier released an upgraded version of the Gemini 2.5 Pro, which showcases enhanced coding functions and has taken the lead on both the WebDev Arena and LMArena leaderboards.
The introduction of the Deep Think mode represents a further step in advancing the AI model. This new reasoning capability enables Gemini 2.5 Pro to evaluate multiple hypotheses prior to formulating a response. Google emphasized that this mode employs a novel research technique distinct from the Thinking versions used in earlier models.
Internal testing has revealed impressive benchmark scores for the Deep Think mode across various metrics. The Gemini 2.5 Pro Deep Think reportedly achieved a score of 49.4 percent on the demanding 2025 UAMO mathematics benchmark; it also performed well on LiveCodeBench v6 and MMMU tests.
Currently, Deep Think is undergoing testing, with Google conducting safety assessments and collaborating with safety experts for feedback. At this time, only a select group of trusted testers can access the reasoning mode via the Gemini API, and no official release date has been announced.
In addition to the Deep Think mode, Google disclosed improvements to the Gemini 2.5 Flash model, which was launched just a month prior. The company highlighted enhancements in key benchmarks related to reasoning, multimodality, code, and extended context. Notably, this updated model operates more efficiently, utilizing 20 to 30 percent fewer tokens.
This upgraded version of Gemini 2.5 Flash is currently available for preview to developers through Google AI Studio. Enterprises can access it on the Vertex AI platform, while individuals can use it via the Gemini app. A wider launch for production use is anticipated in June.
Developers utilizing the Live API will soon gain access to a new feature introduced with the Gemini 2.5 models. The Native Audio Output preview is capable of generating speech that is more expressive and human-like. Users will have the ability to customize various elements of the speech, including tone, accent, and style.
The initial version of this feature comprises three main components. The first, Affective Dialogue, enables the AI to detect emotions expressed in the user’s voice and respond in an appropriate manner. The second component, Proactive Audio, allows the model to disregard background conversations and respond only when addressed directly. Lastly, the Thinking feature leverages Gemini’s reasoning capabilities to verbally tackle complex inquiries.
Moreover, the Gemini 2.5 Pro and Flash models available in the Gemini API and Vertex AI will now include thought summaries, providing insight into the model’s underlying thought processes, which were previously available only in Gemini’s reasoning models. This enhancement allows for a detailed overview that features headers, key points, and information about model actions with each response.
Looking ahead, developers will be able to utilize thinking budgets with the Gemini 2.5 Pro, granting them the ability to manage token consumption before the model generates a response. Finally, the Computer Use agentic function from Project Mariner will soon be integrated into the API and Vertex AI.
