Google's "Deep Think", officially open to the public

Google recently announced that its latest AI feature Deep Think has been formally submitted to the Committee on the Elimination of Discrimination against Women Google AI Ultra Subscribers are open. The move signals that Google is exploring a different development path - "deep thinking" - than the current mainstream speed-seeking AI models. Meanwhile, a more powerful Gemini 2.5 Deep Think The full version has also been made available to selected mathematicians for academic research.

This launch to the public Deep Think version, an improved version based on feedback from early testers and research breakthroughs. It is different from that model that achieved gold standard in this year's International Mathematical Olympiad (IMO) competition. Whereas the competition model takes hours to solve complex math problems, the public version ramps up reasoning dramatically while maintaining daily usability, and its performance still achieves bronze medal levels on the internally evaluated 2025 IMO benchmark.

What is "slow thinking"?

The current AI field is generally looking for faster response times, and the Deep Think The core mechanism of the program does the opposite: it introduces the concept of "extended thinking time".

Deep Think Parallel thinking techniques are used. When faced with a complex problem, the model generates multiple ideas and solution paths simultaneously, and evaluates, revises, and even combines these paths to ultimately select an optimal answer. This process mimics the way humans think when solving a problem: taking the time to explore different perspectives, weighing various potential options, and ultimately forming a well-thought-out conclusion.

In order for the model to effectively utilize the extended inference time, Google has also developed new reinforcement learning techniques that motivate the model to explore more complex inference paths. This allows Deep Think The ability to become a more adept problem solver over time.

Google has moved the Deep Think Offerings to mathematicians Michel van Garrel and other professionals for testing mathematical conjectures.

What are the benefits of Deep Think?

Extended thinking time allows Deep Think Excels at handling tasks that require creativity, strategic planning and incremental improvements.

• Iterative development and design: In tasks that require incremental construction of complex things, theDeep Think demonstrates powerful capabilities. For example, in web development tasks, it is possible to optimize both the functionality of the code and the aesthetic design of the page. The following figure illustrates how to optimize the functionality of code and the aesthetics of a page from the Gemini 2.5 Flash until (a time) Pro re-enter Deep Think Differences in detail, complexity and creativity in generating voxel artwork progressively increase.

  

• Science and Math Explorations: Deep Think The ability to reason about highly complex problems has the potential to be a powerful tool for researchers. It can help construct and explore mathematical conjectures or analyze complex scientific literature in depth.
• Algorithms and Programming: The model is particularly good at solving difficult programming problems. In these problems, it is critical to clearly define the problem, weigh the pros and cons of different options, and consider time complexity.

Its performance is also reflected in several challenging industry benchmarks. Without the use of external tools, theGemini 2.5 Deep Think exist LiveCodeBench V6(a measure of competition programming ability) and Humanity’s Last Exam(measuring expertise in areas such as science and math) and other benchmarks have achieved the most current state-of-the-art performance.

Security and trade-offs

As modeling capabilities increase, security becomes more important. google in Deep Think Security measures are built in throughout the training and deployment cycle of the Tests have shown that with Gemini 2.5 Pro Compare.Deep Think Improvements in content security and tone objectivity, but also a higher tendency to "reject harmless requests". This reflects a trade-off faced by current frontier models in improving security capabilities.

The results of a more detailed security assessment have been published in Deep Think in the model card.

How to use Deep Think

Google AI Ultra Subscribers can now subscribe to the Gemini The application uses the Deep Think, with a fixed limit on the number of times it can be used per day. Users can select from the model drop-down menu the 2.5 ProThen turn on the "Deep Think" switch to activate it. This feature automatically invokes tools such as code execution and Google search, and generates longer responses.

Google is also planning to make its way in the coming weeks through the Gemini API Available to select developers and enterprise testers with and without tools Deep Think version to evaluate its usability in different scenarios.