Artificial Intelligence upgrade, Gemini 2.5 Deep Think, now accessible for Google AI Ultra users
In a significant breakthrough, Google's Gemini 2.5 Deep Think model has topped the charts on LiveCodeBench V6, a competitive coding benchmark, outperforming OpenAI's top models, including OpenAI's o3.
The advanced version of Gemini 2.5 Deep Think was unveiled by Google at their annual developer conference, Google I/O, in May. This new model, designed for everyday tasks, is now part of the Google AI Ultra plan.
Google's AI Premium plan, now renamed Google AI Pro, costs $20 per year. On the other hand, the Google AI Ultra plan, which includes the advanced version of Gemini 2.5 Deep Think, is priced at $250 per month and offers early access to experimental video models, higher limits in tools like NotebookLM, a prototype AI agent manager, and 30TB of storage.
Gemini 2.5 Deep Think is a multi-agent reasoning model that simultaneously tests multiple hypotheses in parallel before choosing the best answer. This novel approach, enabled by greater computational resources and reinforced by new reinforcement learning techniques, powers its superior performance on challenging benchmarks like the International Mathematical Olympiad (IMO).
In the 2025 U.S. Math Olympiad, Gemini 2.5 Deep Think scored big, and an advanced variant of the model even earned a gold medal in internal IMO tests, though the consumer-optimized version reaches a "Bronze-level" IMO performance.
Functionally, Gemini 2.5 Deep Think integrates with tools like code execution and Google Search to enhance its capabilities, focusing on complex problem-solving, algorithmic development, scientific literature navigation, and mathematical conjecture formulation. However, the model’s advanced multi-agent reasoning does entail higher computational costs, limiting access to higher-tier subscription plans such as Google AI Ultra.
Before being rolled out to more users, Deep Think underwent safety checks and received expert input. Now, it's available to Google AI Ultra subscribers, giving them tools to tackle tough problems that require strategy, creativity, and deeper thinking.
In comparison to OpenAI's current models, Gemini 2.5 Deep Think's multi-agent reasoning system tackles questions by exploring multiple ideas in parallel, resulting in improved accuracy and depth over traditional single-agent models like OpenAI's GPT-series. While OpenAI is expected to release GPT-5, its comparative performance against Gemini 2.5 Deep Think in these challenging benchmarks is yet to be seen.
In addition, Gemini 2.5 Deep Think has shown impressive results on Humanity's Last Exam, a tough test covering advanced science and math. With its innovative approach and integration with specialized tools, Gemini 2.5 Deep Think is set to redefine the landscape of AI-assisted problem-solving.
Artificial-intelligence advancements in Google's Gemini 2.5 Deep Think model have set new standards, outperforming OpenAI's GPT series on complex problem-solving tasks, as demonstrated in LiveCodeBench V6 and the International Mathematical Olympiad (IMO). Its multi-agent reasoning system explores multiple ideas in parallel, enhancing its accuracy and depth over traditional single-agent models.
With its superior performance and integration with specialized tools like code execution and Google Search, Gemini 2.5 Deep Think under the Google AI Ultra plan has become a powerful instrument for tackling tough problems that require strategy, creativity, and deeper thinking, breaking new grounds in AI-assisted problem-solving.
