Page Loader
Google unveils 'Deep Think' to supercharge flagship Gemini AI model
Deep Think weighs multiple answers before responding

Google unveils 'Deep Think' to supercharge flagship Gemini AI model

May 21, 2025
12:19 am

What's the story

Google has unveiled Deep Think, a new reasoning mode for its cutting-edge Gemini 2.5 Pro AI model. The feature, announced at the Google I/O 2025 event, aims to improve the model's performance by weighing multiple answers before responding. Demis Hassabis, head of Google DeepMind, called it a technology that "pushes model performance to its limits" with cutting-edge research in thinking and reasoning techniques.

Benchmark success

Deep Think boosts Gemini 2.5 Pro's performance

Though Google hasn't revealed exact details of Deep Think's working, it is speculated to be similar to OpenAI's o1-pro and upcoming o3-pro models. They are believed to use an engine to search for and synthesize the best solutions to problems. As per Google, Deep Think has allowed Gemini 2.5 Pro to outperform competitors in multiple benchmarks including LiveCodeBench and MMMU tests for coding evaluation, perception, and reasoning skills.

Limited access

Deep Think's rollout and safety evaluations

Deep Think is now available for "trusted testers" through the Gemini API. Google has said it's taking more time for safety assessments before a broader rollout of Deep Think. This cautious approach appears to be a testament to the company's commitment to ensuring the safety and reliability of its AI technologies before making them available to a wider audience.

Model upgrade

Google updates Gemini 2.5 Flash

Along with Deep Think, Google has also refreshed its budget-friendly Gemini 2.5 Flash model. The new iteration is tuned to deliver better performance on tasks related to coding, reasoning, multimodality, and long context. It is now more efficient than its predecessor and can be previewed in Google's AI Studio as well as Vertex AI platforms and the company's Gemini apps.

Others

Google introduces Gemini Diffusion

Google has also unveiled a new model, Gemini Diffusion, which it touts as "very fast," providing output 4-5 times faster than similar models. The speed is comparable to that of models twice its size. Gemini Diffusion is available to "trusted testers" starting today, further extending Google's portfolio of cutting-edge AI technologies for different use cases.