Overview
- Gemini 2.5 Deep Think mode is now live for Google AI Ultra subscribers in the Gemini app, offering a limited number of advanced prompts per day.
- Deep Think employs parallel agent processes and extended inference time to explore and refine multiple hypotheses before delivering solutions.
- The model’s coding performance has risen to an 87.6% score on LiveCodeBench and it achieves bronze-level results on the 2025 International Mathematical Olympiad benchmark.
- Deep Think automatically integrates with tools such as code execution and Google Search to support longer, more detailed responses for coding, scientific and research tasks.
- In the coming weeks, Google will provide with-and-without-tools versions of Deep Think via the Gemini API to select enterprise and developer testers.