Google Unveils Gemini 2.5 Deep Think for AI Ultra Subscribers

Google has just released its most powerful Gemini model to date, Gemini 2.5 Deep Think, but access is restricted to subscribers of the $250 AI Ultra plan. This model, which was first announced at the I/O conference in May, is designed to handle complex queries that require more compute resources than other models. Here’s a deeper dive into what makes Gemini 2.5 Deep Think stand out and why it matters for practitioners.

What Changed Technically?

• Enhanced Thinking Time: Deep Think builds on the foundation of Gemini 2.5 Pro but significantly increases the "thinking time" with greater parallel analysis. This means the model explores multiple approaches to a problem, revisiting and remixing hypotheses to generate higher-quality outputs.
• Increased Compute Resources: Due to its advanced capabilities, Deep Think requires more computational power. As a result, it can take several minutes to produce answers, which is particularly beneficial for tasks that demand deep reasoning.

Why It Matters

• Complex Query Handling: Deep Think excels in handling complex queries across various domains, including design aesthetics, scientific reasoning, and coding. This makes it a valuable tool for professionals who need AI assistance with intricate problems.
• Benchmark Performance: Google has benchmarked Deep Think against other models, showing significant improvements. For instance, it outperforms the standard Gemini 2.5 Pro and competing models like OpenAI o3 and Grok 4 in the Humanity’s Last Exam (HLE) benchmark. HLE consists of 2,500 complex, multi-modal questions covering over 100 subjects. While other models score around 20-25%, Deep Think achieves a remarkable 34.8%.
• Mathematical Proficiency: Mathematics is a key focus for Deep Think. It demonstrates strong performance in the AIME benchmark and has even achieved bronze medal status in the 2025 International Mathematical Olympiad (IMO) test. Google used a specially trained version of Deep Think to earn an IMO gold medal, although this version is currently limited to trusted testers.

Implementation Details

• Parallel Analysis: Deep Think employs parallel analysis to explore multiple hypotheses simultaneously. This approach allows it to revisit and refine its solutions, leading to more accurate and nuanced outputs.
• Compute Requirements: The increased compute requirements mean that Deep Think takes several minutes to generate answers. While this might be a drawback for real-time applications, it’s highly beneficial for tasks that benefit from deep, multi-faceted analysis.
• Access and Availability: Currently, Deep Think is only available to Google AI Ultra subscribers via the Gemini app and web interface. It’s not included in the main model menu but can be accessed as a tool alongside others like Deep Research and Canvas.

Future Prospects

Google plans to release the IMO gold medal-winning version of Deep Think more widely in the future, although it is currently restricted to trusted testers. This advanced version, which can run for hours to solve complex problems, could further expand the capabilities of AI in specialized domains.

Conclusion

Gemini 2.5 Deep Think represents a significant step forward in AI research and application, particularly for handling complex queries that require deep reasoning and analysis. While its compute-intensive nature limits its availability to high-tier subscribers, it offers promising potential for professionals and researchers who need powerful AI tools.