Google is significantly advancing its Gemini 2.5 artificial intelligence models, highlighted by the introduction of “Deep Think,” an experimental reasoning mode for its flagship Gemini 2.5 Pro. Revealed at Google I/O 2025, Deep Think aims to tackle highly complex math and coding by enabling the model to consider multiple hypotheses before responding. This development, coupled with performance enhancements for the speed-oriented Gemini 2.5 Flash and new capabilities across both models, including native audio output, underscores Google’s commitment to providing more powerful and versatile AI tools.
The “Deep Think” feature, as detailed by Google DeepMind, is designed to elevate Gemini 2.5 Pro’s analytical capabilities. Demis Hassabis, head of Google DeepMind, stated during a press briefing that Deep Think “pushes model performance to its limits,” utilizing cutting-edge research. This approach might be similar to techniques used by competitors like OpenAI.
Early results show Deep Think enabling Gemini 2.5 Pro to top the demanding LiveCodeBench coding evaluation and outperform OpenAI’s o3 on the MMMU multimodal skills test. However, Google is taking a cautious approach, initially releasing Deep Think to trusted testers via the Gemini API for further safety evaluations, with availability on Vertex AI for these testers to follow soon.
Alongside these high-end enhancements, Gemini 2.5 Flash, optimized for speed and efficiency, also received a substantial update. The model now boasts improved performance in reasoning, multimodality, coding, and long-context tasks, while being 20-30% more token-efficient, as detailed by Google.
The updated 2.5 Flash is currently in preview within Google AI Studio, Vertex AI, and the Gemini app for consumers, with general availability for production use expected in early June.
This builds upon its April introduction, where it was presented as a solution for high-frequency tasks featuring “dynamic and controllable computing.” Geotab, a fleet management company, noted in a statement shared by Google that for their analytics agent, Gemini 2.5 Flash on Vertex AI offers an “excellent balance” and “good consistency,” notably “delivering 25% faster response times on subjects where it has less familiarity.”
Enhanced Interaction And Broader Capabilities
The Gemini 2.5 series now incorporates native audio output via the Live API, enabling more natural and expressive AI conversations. This includes audio-visual input and the ability for users to direct the AI’s tone and accent, and even instruct it to use tools like search. Experimental Live API features include Affective Dialogue, allowing the model to detect and respond to user emotions, and Proactive Audio for discerning response cues in noisy environments.
Furthermore, new text-to-speech previews for both models support multiple speakers and over 24 languages, with the ability to capture subtle nuances like whispers.
Security across the Gemini family has been bolstered with advanced safeguards, according to Google DeepMind, against threats such as indirect prompt injections. Looking forward, Project Mariner, an experimental AI agent functioning as a Chrome extension to autonomously browse and interact with websites, will be integrated into the Gemini API and Vertex AI.
Developer experiments for Project Mariner are slated for this summer, with initial access for Google AI Ultra subscribers in the US. Gemini Live, featuring camera and screen sharing powered by Project Astra, is also now rolling out to all users on iOS.
Developer Focus And Evolving Ecosystem
Google continues to refine the developer experience by introducing thought summaries in the Gemini API and Vertex AI to provide clearer insights into model reasoning. Thinking budgets, which allow for control over computational costs, are being extended from Flash to 2.5 Pro. These updates are part of a rapid iteration cycle for Gemini. An ‘I/O Edition’ of Gemini 2.5 Pro was already released on May 6, focusing on better coding skills.
This rapid deployment has, at times, drawn scrutiny. The initial model card for Gemini 2.5 Pro, published weeks after its public availability in late March, was termed “meager” by Kevin Bankston of the Center for Democracy and Technology. This fueled concerns about transparency in AI safety as models are quickly brought to market.
The Gemini AI chatbot app now serves over 400 million active monthly users, a figure announced by CEO Sundar Pichai at I/O 2025.
New Subscription Tiers And Integrations
Coinciding with these model enhancements, Google introduced a new “Google AI Ultra” subscription tier at $249.99/month, which will offer early and maximum access to features like Gemini 2.5 Pro with Deep Think for consumers, Veo 3 (a new video generation model with native audio), and Project Mariner. The existing “AI Premium” tier is being rebranded to “Google AI Pro” at $19.99/month.
Gemini is also being integrated more deeply into other Google products. Gemini will be available in the Chrome browser on desktop for AI Pro and Ultra subscribers to clarify or summarize web content. Additionally, Firebase Studio‘s App Prototyping agent now utilizes the latest Gemini 2.5 models and new API features like native audio output. These moves illustrate Google’s strategy to embed its increasingly sophisticated AI across its ecosystem, aiming to make its tools more intelligent and interactive.
