Recently, Google DeepMind announced that its most powerful AI model, Gemini 2.5 Deep Think, is now available to Google AI Ultra subscription users. This model not only won a gold medal in the 2025 International Mathematical Olympiad (IMO) competition, but also demonstrated impressive performance across multiple fields with its innovative "parallel thinking" and reinforcement learning technologies.

 Gemini 2.5 Deep Think: The New Peak of AI Reasoning

Gemini 2.5 Deep Think is the most advanced model in the Gemini 2.5 series, designed to handle complex tasks. Its core highlights include the introduction of "Parallel Thinking" and a new reinforcement learning technology, allowing the model to simulate the process of human brainstorming, explore multiple reasoning paths simultaneously, and compare them to generate more accurate and creative answers. Compared to traditional AI models' linear reasoning, this capability makes Deep Think particularly outstanding in solving complex problems.

QQ20250804-110503.jpg

Main technological breakthroughs:

1. Parallel Thinking Mechanism: Deep Think uses a multi-agent system, allowing multiple AI "agents" to work on a problem at the same time, exploring different hypotheses and integrating results. This approach not only enhances the depth of reasoning but also significantly improves the ability to solve complex tasks such as mathematics, science, and coding.

2. Reinforcement Learning Optimization: Google developed a new reinforcement learning technique that encourages the model to continuously optimize its strategies during the reasoning process. This makes Deep Think more efficient when handling tasks that require gradual improvement, such as algorithm design and strategic planning.

3. Multimodal and Long Context Support: Gemini 2.5 Deep Think supports text, audio, images, and video input, and has a context window of 1 million tokens, enabling it to handle large datasets and suitable for various scenarios from academic research to real-time applications.

 IMO Gold Medal Certification: A Milestone in Mathematics and Reasoning

In the 2025 International Mathematical Olympiad (IMO) competition, an optimized version of Gemini 2.5 Deep Think achieved a score of 35 out of 42, winning a gold medal, demonstrating its top-level strength in mathematical reasoning. According to Professor Gregor Dolinar, the president of IMO, Deep Think's solutions were "clear, precise, and in many cases easier to understand than those of human participants."

Breakthroughs in Mathematics and Science:

- Deep Think successfully solved five out of six problems in the IMO competition, proving its exceptional ability in tackling complex mathematical problems.

- Compared to last year's AlphaProof and AlphaGeometry2 models (which won silver), Deep Think uses pure natural language processing, eliminating dependence on specific domain languages, making its reasoning process more general and flexible.

- The public version of Deep Think, optimized for daily use, can still achieve bronze level performance in the IMO benchmark test, balancing performance with practicality.

 Outstanding Performance in Benchmark Tests: Coding and Cross-Domain Knowledge

Gemini 2.5 Deep Think has shown excellent performance in multiple authoritative benchmark tests, solidifying its leading position in the AI field:

- LiveCodeBench V6: In this competitive coding benchmark test, Deep Think achieved a score of 87.6%, surpassing xAI's Grok4 (79%) and OpenAI's o3 (72%), showcasing its strong capabilities in handling complex programming tasks.

- Humanity’s Last Exam (HLE): This comprehensive test covering mathematics, science, and humanities includes approximately 3,000 expert-level questions. Deep Think led with a score of 34.8%, far ahead of Grok4 (25.4%) and o3 (20.3%).

- WebDev Arena and LMArena: Deep Think performed exceptionally well in web development and learning assistance areas, becoming a leader in relevant rankings.

These results show that Deep Think is not only good at mathematics and coding, but also capable of handling complex knowledge problems across multiple domains, providing researchers and developers with powerful tools.

 User Accessibility: Limited to AI Ultra Subscription Users

Gemini 2.5 Deep Think is now available through the Gemini mobile app (iOS and Android) to Google AI Ultra plan subscribers, with a subscription fee of $249.99 per month (new users get a discounted rate of $124.99 for the first three months). Users can use a fixed number of Deep Think prompts daily, and the model automatically integrates tools such as code execution and Google search to generate more detailed responses.

Google also plans to offer Deep Think versions with and without tools to trusted testers, including mathematicians and developers, in the coming weeks via the Gemini API, further exploring its potential applications in enterprise and development scenarios.

 Industry Impact and Future Outlook

The release of Gemini 2.5 Deep Think marks another leap in AI reasoning capabilities. The application of parallel thinking and reinforcement learning technologies not only enhances the model's performance in academic and coding tasks but also opens up new possibilities for creative tasks such as design optimization and strategic planning. Google DeepMind stated that Deep Think will continue to evolve in the future, aiming to achieve a perfect score in the IMO and expand into more fields.

AIbase Perspective: The launch of Gemini 2.5 Deep Think indicates that the AI industry is moving from simple pattern recognition toward deeper reasoning and creativity. However, high subscription fees and computational resource demands may limit its accessibility. In the future, how Google balances performance, cost, and accessibility will determine whether Deep Think can truly become a "game-changer" in the AI field.

Conclusion

Gemini 2.5 Deep Think has set a new benchmark for the future development of AI with its gold medal performance in the IMO and cross-domain capabilities.