Google has officially rolled out Deep Think, an advanced version of its Gemini 2.5 model, to Google AI Ultra subscribers via the Gemini app. This release brings powerful creative problem-solving capabilities, especially for developers, mathematicians, and researchers working on complex, multi-step tasks.

What is Deep Think?

Deep Think is a specialized variation of Gemini 2.5, tuned for extended reasoning and creative analysis. It’s designed to think more like a human—exploring multiple solutions in parallel, refining answers over time, and delivering results that are both thoughtful and accurate.

This version integrates feedback from trusted testers and breakthroughs in reinforcement learning, resulting in faster responses and improved usability without sacrificing deep reasoning. Internally, the model achieved bronze-level performance on the 2025 International Mathematical Olympiad (IMO) benchmark and is a lighter, faster counterpart to the full model that secured gold-standard results in the actual IMO.

How It Works

Deep Think extends the model’s “thinking time,” allowing it to explore more hypotheses in parallel before finalizing an answer. This is particularly useful in areas such as:

• Mathematical and scientific reasoning

• Strategic algorithm development

• Iterative web and UI design

• Advanced coding and debugging

Google says Deep Think now leads in benchmarks like LiveCodeBench V6 and Humanity’s Last Exam, outperforming models like Gemini 2.5 Pro, OpenAI’s GPT-4 (o3), and Grok 4.

Responsible AI by Design

The Deep Think model includes several safety advancements. According to Google’s internal evaluations, it shows improved content safety and tone objectivity. However, it also tends to be more cautious—sometimes refusing even harmless prompts to reduce risk.

Details on responsible development practices and safety mitigations are available in its official model card.

How to Use Deep Think

Google AI Ultra subscribers can access Deep Think in the Gemini app by selecting the 2.5 Pro model and toggling the Deep Think option in the prompt bar. It supports tools like code execution and Google Search, and is capable of producing significantly longer, more detailed responses.

Additionally, Google plans to open up API access to trusted testers, including developers and enterprise users, in the coming weeks.

This release marks a major milestone in Google’s mission to build capable, creative, and responsible AI tools for the future of development and research.

Explore the full announcement on Google’s official blog