Deep Think Improves Gemini AI Performance - Google AI Updates

Google Enhances Gemini AI Models with New Capabilities

Google is implementing upgrades to its most advanced Gemini AI models. These enhancements aim to improve performance and broaden the application of its AI technology.

Introducing Deep Think for Gemini 2.5 Pro

At Google I/O 2025, the company unveiled Deep Think, an advanced reasoning mode designed for the Gemini 2.5 Pro model. This new mode allows the AI to evaluate multiple potential responses before formulating an answer.

According to Demis Hassabis, head of Google DeepMind, Deep Think represents a significant leap forward. It leverages the latest research in thinking and reasoning, incorporating parallel processing techniques to maximize model performance.

While the specific mechanisms behind Deep Think remain somewhat undisclosed, it shares similarities with OpenAI’s o1-pro and the forthcoming o3-pro models. These models utilize a search engine to identify and synthesize optimal solutions to complex problems.

Performance Benchmarks and Availability

Google reports that Deep Think has enabled Gemini 2.5 Pro to achieve top scores on LiveCodeBench, a demanding coding evaluation benchmark. Furthermore, it outperformed OpenAI’s o3 on MMMU, a test assessing perception and reasoning abilities.

Currently, Deep Think is accessible to a select group of “trusted testers” through the Gemini API. Google is prioritizing thorough safety evaluations before a wider public release.

Updates to Gemini 2.5 Flash

In addition to Deep Think, Google has released an update for its more affordable Gemini 2.5 Flash model. This update enhances the model’s capabilities in areas such as coding, multimodal processing, reasoning, and handling extensive context.

The updated 2.5 Flash model is also more efficient than its predecessor. It is now available for preview on Google’s AI Studio and Vertex AI platforms, as well as within the company’s Gemini applications.

Developers can expect general availability of the improved Gemini 2.5 Flash sometime in June.

Introducing Gemini Diffusion

Google is also introducing Gemini Diffusion, a new model characterized by its speed. The company claims it delivers outputs 4-5 times faster than comparable models, while achieving performance levels comparable to models twice its size.

Gemini Diffusion is currently being tested by “trusted testers” and became available today.