Google says Gemini 2.5 models are only getting better with Deep Think and Flash
What you need to know
- Google’s Gemini 2.5 model picks up 2.5 Flash to be even faster and more efficient with its actions.
- Deep Think, Google’s new reasoning mode, sets its sights on Gemini 2.5 Pro; however, there’s an abundance of testing required before it arrives.
- Gemini 2.5 Pro gains native audio output controls, meaning developers can alter the AI’s tone, accent, and “style of speech.”
I/O wouldn’t be the same without a few AI announcements, and that’s exactly what Google is bringing to users.
In a press release, Google detailed how its Gemini 2.5 models, including 2.5 Pro, are going to advance in the near future thanks to a few updates. To begin, Gemini is going to receive Google’s 2.5 Flash model. Not only has the company touted this AI as its “most powerful” version, but its latest updates improve benchmarks for reasoning and multimodality. Moreover, Google says 2.5 Flash is now better (more efficient) at code and long context.
All Gemini users will find 2.5 Flash available in the app. Additionally, Google says this update will hit Vertex AI users with enterprise and for developers in the Google AI Studio.
Gemini’s thinking capabilities are seemingly set for improvements, as Google details the incorporation of a new reasoning mode called “Deep Think.” This mode is said to push Google’s AI into “considering multiple hypotheses” before delivering its response. This is reportedly en route to 2.5 Pro (under the guise of 2.5 Pro Deep Think). This mode is in testing. Google states it is taking more time to conduct “frontier safety evaluations” and to hear more from the experts.
Once these initial hurdles are cleared, Google will prepare to release Deep Think more widely to users.
Gemini 2.5 Rises Higher
Aside from these key highlights, Google details even more Gemini 2.5 capabilities as a result of its updates. There are three major updates, the first bringing “native” audio output and improvements to the Live API. Appearing today (May 20), Live API will deliver a preview of its audio-visual input and native audio out dialogue. This will give users conversational building capabilities within Gemini.
Moreover, this also gives users the chance to customize how Gemini speaks by altering its tone, accent, and “style of speaking.”
This audio update brings an experiment for a trio of early features. Users will have a chance to play with “Affective Dialogue,” which lets the AI “detect” emotions in a user’s voice to facilitate an “appropriate” response. The Proactive Audio test will let Gemini ignore voices in your background, as it awaits a more proper time to respond. Thinking hits Live API for Gemini, improving its complex task capabilities.
Project Mariner’s computer use functionality is headed for Gemini and Vertex AI. Additionally, Google states Gemini 2.5 will be receiving another security boost. This time, the AI model is picking up enhanced protections for maliciously embedded instructions and “indirect prompt injection” attacks.
Developers get a little something
With so much extra thought (thinking) power headed for Gemini and Vertex AI, Google is giving developers insightful summaries to help. The summaries are said to organize the AI’s thinking process, as well as offer information about its actions. The entire purpose is to help developers debug the AI in case there are any lingering issues.
Cost control (via thinking budget) is on its way to Gemini 2.5 Pro “in the coming weeks.” This will be joined by Google’s “generally available model,” too.
Lastly, Gemini 2.5 is bringing MCP (Model Context Protocol) support. This should give developers an easier time when integrating open-source tools into their Gemini work. Google teases it’s also looking into MCP servers and more “hosted” tools.
Post Comment