At the Google I/O Developer Conference on May 20, 2025, Google DeepMind officially launched Gemini2.5Pro with the Deep Think mode, an experimental advanced reasoning mode that sets a new benchmark for AI in handling complex tasks. The Deep Think mode enhances Gemini2.5Pro's performance in mathematics, coding, and multi-modal reasoning to industry-leading levels through parallel reasoning technology.
Parallel Reasoning Technology Significantly Enhances Performance on Complex Tasks
The Deep Think mode employs cutting-edge parallel reasoning technology, enabling the model to explore multiple hypothetical paths before generating responses, thereby improving the accuracy and depth of answers. Google stated that this mode achieved excellent results in the 2025 USAMO (USA Mathematical Olympiad) test, topped the competitive programming benchmark LiveCodeBench, and scored an impressive 84% on the multi-modal reasoning test MMU. These results indicate that the Deep Think mode is particularly adept at solving high-difficulty mathematical and coding problems requiring multi-step reasoning, rivaling or even surpassing competitors like OpenAI’s o1-pro.
Modular Control and Developer-Friendly Design
The Deep Think mode introduces the **"thinking budget" feature**, allowing developers to flexibly balance response quality, latency, and cost by adjusting the number of tokens used during reasoning. This enables Gemini2.5Pro to adapt to various scenarios, from rapid prototyping to deep analysis. Additionally, Google has added the **"thought summaries"** feature to the Gemini API and Vertex AI, presenting the model's reasoning process in a clear format to facilitate debugging and optimization of interactive experiences for developers.
Initially Available to Trusted Testers, Prioritizing Safety
Currently, the Deep Think mode is only available via the Gemini API to trusted testers, with Google emphasizing additional safety assessments to ensure robustness before widespread application. Developers are highly anticipating the reasoning capabilities of this mode, but also note that its practical performance requires further validation. Google plans to gradually expand its availability through Google AI Studio and Vertex AI over the next few weeks, with expectations for broader deployment by early June.
The release of the Gemini2.5Pro Deep Think mode marks a major breakthrough for Google in the field of AI reasoning, offering developers more efficient tools through its parallel reasoning and modular control design. Compared to traditional single reasoning paths, the innovative approach of the Deep Think mode is expected to spark a new trend in academic research, software development, and data analysis. AIbase will continue to track the actual application effects and user feedback of this mode, looking forward to its performance in pushing the boundaries of AI technology.