Google is upgrading its most capable Gemini AI models.
On Tuesday at Google I/O 2025, the company announced Deep Think, an “enhanced” reasoning mode for its flagship Gemini 2.5 Pro model. Deep Think allows the model to consider multiple answers to questions before responding, boosting its performance on certain benchmarks.
“[Deep Think] pushes model performance to its limits,” said Demis Hassabis, head of Google DeepMind, Google’s AI R&D org, during a press briefing. “It uses our latest cutting-edge research in thinking and reasoning, including parallel techniques.”
Google was vague on the inner workings of Deep Think, but the technology could be similar to OpenAI’s o1-pro and upcoming o3-pro models, which likely use an engine to search for and synthesize the best solution to a given problem.
Google says that Deep Think enabled Gemini 2.5 Pro to top LiveCodeBench, a challenging coding evaluation. Gemini 2.5 Pro Deep Think also beat OpenAI’s o3 on MMMU, a test for skills like perception and reasoning.
Deep Think is available to “trusted testers” via the Gemini API as of this week. Google said that it’s taking additional time to conduct safety evaluations before rolling out Deep Think widely.
Alongside Deep Think, Google has introduced an update to its budget-oriented Gemini 2.5 Flash model that allows the model to perform better on tasks involving coding, multimodality, reasoning, and long context. The new 2.5 Flash, which is also more efficient than the version it replaces, is available for preview in Google’s AI Studio and Vertex AI platforms as well as the company’s Gemini apps.
Google says that the improved Gemini 2.5 Flash will become generally available for developers sometime in June.
Lastly, Google is introducing a model called Gemini Diffusion, which the company claims is “very fast” — delivering output 4-5 times quicker than comparable models and rivaling the performance of models twice its size. Gemini Diffusion is available beginning today to “trusted testers.”
