Google just dropped one of its most ambitious AI upgrades yet. The company is rolling out Deep Think, a powerful new mode inside the Gemini app that boosts the reasoning power of its top-tier model, Gemini 2.5 Ultra.
Starting today, Deep Think is available to Google AI Ultra subscribers, with a broader API release for trusted testers expected in the coming weeks.
What is Deep Think?
Google says Deep Think excels at challenges that require multi-step reasoning, strategy and iteration. Think of it as giving Gemini more thinking time before the AI responds.
Now, instead of rushing to deliver an answer, Gemini engages in parallel thinking, a process that allows it to generate multiple ideas, evaluate them simultaneously and refine its approach before responding.
It's a move that brings Gemini one step closer to emulating human-style problem-solving, whether it's debugging code, website building or tackling advanced math.
Internally, the model has reached bronze-level performance on the 2025 International Mathematical Olympiad (IMO) benchmark and outperformed competitors on difficult reasoning tests like LiveCodeBench V6 and Humanity’s Last Exam, which measure capabilities in science, coding and complex logic.
The company also tested Deep Think with academic researchers, including mathematician Michel van Garrel, to explore its potential for tackling theoretical problems and advancing scientific discovery.
While the original research model behind Deep Think took hours to solve Olympiad-level math problems, the version launching today is optimized for faster, more practical day-to-day use.
How to try it
To try it out, Ultra subscribers can simply toggle Deep Think on in the Gemini app when using Gemini 2.5 Pro.
The mode supports longer, more detailed responses and works with tools like Google Search and code execution, though usage is currently limited to a fixed number of prompts per day.
Prioritizing safety
As with previous Gemini updates, Google is placing a strong emphasis on safety.
The company reports that Deep Think shows improved tone and content moderation over Gemini 2.5 Pro, though it may sometimes refuse benign requests.
Google says it’s actively performing frontier safety evaluations and working on mitigations as model capabilities advance.
Bottom line
For Ultra subscribers, this new feature should support app building, exploring a scientific theory and many more tasks that require a more thoughtful response.
While Deep Think is currently limited to $249/month Gemini Ultra subscribers, there’s a good chance it could become more widely accessible in the future. Google has a history of offering some of its most powerful AI features, like Gemini Pro and image generation, to free-tier users after initial testing.
If Deep Think proves successful and stable, it’s likely that Google will broaden access to showcase its capabilities and stay competitive alongside other tech giants.
Follow Tom's Guide on Google News to get our up-to-date news, how-tos, and reviews in your feeds. Make sure to click the Follow button.
