Google officially unveiled “Deep Think” at Google I/O 2025, a powerful new reasoning mode integrated into its flagship Gemini 2.5 Pro model. By enabling the AI to evaluate multiple potential solutions before generating a final response, Google is significantly sharpening the model’s performance on complex, logic-heavy benchmarks.
Pushing AI Reasoning to the Limits
Demis Hassabis, head of Google DeepMind, described the technology as a breakthrough that pushes model performance to its absolute limits. “It uses our latest cutting-edge research in thinking and reasoning, including parallel techniques,” Hassabis noted during a recent press briefing.
While Google remains guarded regarding the specific architecture of Deep Think, industry analysts suggest the mechanism mirrors the logic employed by OpenAI’s o1-pro and upcoming o3-pro models. These systems typically utilize an internal engine to synthesize the most accurate solution from a range of possibilities, effectively “thinking” before they speak.
Benchmark Dominance
The impact of this upgrade is already visible in rigorous testing environments. Google reports that Gemini 2.5 Pro with Deep Think has secured the top spot on LiveCodeBench, a highly demanding coding evaluation. Furthermore, it has outperformed OpenAI’s o3 model on the MMMU benchmark, which tests advanced perception and reasoning capabilities.
Currently, Deep Think is accessible only to a select group of “trusted testers” via the Gemini API. Google has confirmed it is prioritizing comprehensive safety evaluations before a broader public rollout.
Gemini 2.5 Flash and Diffusion Updates
Beyond the flagship model, Google announced a significant refresh for its efficient, budget-friendly Gemini 2.5 Flash model. This updated iteration delivers superior performance across coding, multimodality, and long-context tasks while operating with increased efficiency. It is now available for preview in Google AI Studio, Vertex AI, and Gemini apps, with general availability for developers scheduled for June.
Additionally, the company introduced “Gemini Diffusion,” a high-speed model designed for rapid output generation. Google claims this new model is four to five times faster than comparable alternatives, rivaling the performance of models twice its size. Access to Gemini Diffusion is currently limited to trusted testers starting today.