Google’s big I/O 2025 event is underway in California, and it’s all about artificial intelligence, just as we suspected. That is, Gemini is improving across the board, with Google announcing new features for the Gemini app and integrating its chatbot into other products.
Google claims that Gemini 2.5 Pro is the best and fastest AI chatbot in the world, backing that up with benchmark tests that show Google is ahead of the competition. Some of these tests reveal that the second-best AI model after Gemini 2.5 Pro isn’t a new ChatGPT version but Google’s Gemini 2.5 Flash. That’s the lighter, faster version of Gemini 2.5, which is also getting an update on Tuesday.
Google is also adding a few new features to the Gemini app experience, including better voice support, new Deep Think functionality, and an expansion of Project Mariner. The latter is Google’s AI agent that gives the AI “computer use” capabilities, allowing it to browse websites and make purchases on your behalf.
Gemini 2.5 Flash
Google says Gemini 2.5 Flash will be better in “nearly every dimension,” showing improvements in reasoning, coding, and handling long contexts. The lighter model should also be more efficient, using 20–30% fewer tokens in Google’s tests.
The new Gemini 2.5 Flash is rolling out in preview mode in Google AI Studio (developers), Vertex AI (enterprise), and the Gemini app (everyone).
A final version of Gemini 2.5 Flash will be more widely available in June.
Better voice experiences
Gemini is also becoming more natural and expressive in voice chats. Users will be able to customize the AI’s tone, accent, and style to match the type of content it’s handling.
The new native audio output will support the following early features:
Affective Dialogue, where the model detects emotion in the user’s voice and responds appropriately.
Proactive Audio, where the model filters out background conversations and knows when to respond.
Thinking in the Live API, where the model uses Gemini’s reasoning capabilities to support more complex tasks.
The Gemini voice experience will also support tool use, like web searching.
Google also announced new previews for text-to-speech in 2.5 Pro and Flash, offering “first-of-its-kind” support for multiple speakers. Gemini will support text-to-speech with two voices through native audio in over 24 languages. The feature will debut in the Gemini API.
Deep Think
Like ChatGPT, Gemini can offer Deep Research reports on whatever you’re interested in, but Google is going further. At I/O 2025, it announced a new Deep Think functionality, an “enhanced reasoning mode” that lets the AI consider multiple hypotheses before responding to a prompt.
Gemini 2.5 Pro Deep Think earned an “impressive” score on the 2025 USAMO test, one of the hardest math benchmarks. It leads LiveCodeBench and scored 84% on MMMU, a multimodal reasoning test.
Deep Think won’t roll out to Gemini apps until Google completes its safety evaluations for this frontier model. Trusted testers will get access through the Gemini API.
Project Mariner
I haven’t exactly been a big fan of Project Mariner, especially compared to ChatGPT Operator, though I haven’t had access to either. We’ve barely heard anything about Mariner since December, but Google has been testing it with a select group of users.
Project Mariner’s abilities, such as web browsing and making purchases for users, will also be available in Google Search AI Mode.
At I/O 2025, Google said it’s ready to bring Project Mariner’s computer use capabilities to the Gemini API and Vertex AI.
Thought summaries and thinking budgets
Both Gemini 2.5 Pro and 2.5 Flash will include thought summaries in the Gemini API and Vertex AI, letting you see the AI’s raw “thoughts.”
Gemini 2.5 Flash has had a thinking budget feature from the start, allowing developers to control how much the AI thinks (how many tokens it uses) before responding. This feature is now rolling out to Gemini 2.5 Pro as well.
MCP support and better security
Google has added native SDK support for Model Context Protocol (MCP) definitions in the Gemini API, which allows developers to integrate open-source tools with Gemini.
Gemini 2.5 Pro now includes enhanced protections against security threats, such as indirect prompt injections, making it the most secure model family to date.
