Google Gemini AI Gets "Deep Think" Upgrade for Enhanced Reasoning

Google Unveils "Deep Think" for Enhanced Gemini AI Reasoning

Google announced a significant upgrade to its flagship Gemini 2.5 Pro AI model at Google I/O 2025: Deep Think. This new reasoning mode allows the model to evaluate multiple solutions before responding, boosting performance on various benchmarks.

"Deep Think pushes model performance to its limits," said Demis Hassabis, head of Google DeepMind. "It uses our latest cutting-edge research in thinking and reasoning, including parallel techniques."

While details about Deep Think's inner workings remain limited, it appears similar to OpenAI's advanced models, which use sophisticated search and synthesis engines for problem-solving.

Deep Think Drives Gemini 2.5 Pro Performance

Thanks to Deep Think, Gemini 2.5 Pro achieved top scores on LiveCodeBench, a challenging coding evaluation. It also surpassed OpenAI's o3 model on MMMU, a test assessing perception and reasoning skills.

Deep Think is currently available to trusted testers via the Gemini API, with wider availability pending safety evaluations.

Gemini 2.5 Flash and Gemini Diffusion Also Get Upgrades

Google also announced updates to Gemini 2.5 Flash, its budget-friendly model. This updated version offers improved performance in coding, multimodality, reasoning, and long context handling, along with increased efficiency. It's available for preview in Google AI Studio, Vertex AI, and Gemini apps, with general availability expected in June.

Finally, Google introduced Gemini Diffusion, a significantly faster model. The company claims it delivers output 4-5 times quicker than comparable models, rivaling the performance of models twice its size. Gemini Diffusion is currently available to trusted testers.

These advancements highlight Google's ongoing commitment to pushing the boundaries of AI capabilities and performance.