Google is upgrading its most capable Gemini AI models.
On Tuesday at Google I/O 2025, the company announced Deep Think, the “enhanced” reasoning mode for its flagship Gemini 2.5 Pro model. Deep Think allows models to consider multiple answers to questions before responding, and improve performance with specific benchmarks.
“(Deep Thoughts) pushes the performance of a model to the limit,” said Demis Hassabis, head of Google Deepmind, Google’s AI R&D organization, during the press conference. “We use the latest cutting-edge research into thinking and reasoning, including parallel techniques.”
Google has been ambiguous about the inner workings of Deep Think, but this technology may resemble Openai’s O1-Pro and upcoming O3-Pro models. This could potentially use the engine to search and synthesize the best solution for a particular problem.
Google says Deep Think has enabled Top LiveCodebench from the Gemini 2.5 Pro. This is a challenging coding evaluation. Gemini 2.5 Pro Deep Think defeated Openai’s O3 with MMMU. This is a test of skills such as perception and reasoning.
As of this week, DEEP Think is available to “trusted testers” through the Gemini API. Google said it takes additional time to carry out a safety assessment before thinking too deeply.
In addition to Deep Think, Google has introduced budget-oriented Gemini 2.5 Flash model updates, allowing models to improve performance on tasks that involve coding, multimodality, inference, and long contexts. Additionally, the new 2.5 flash, which is more efficient than the version you replace, can be previewed on Google’s AI Studio and Vertex AI platforms, as well as the company’s Gemini app.
Google says the improved Gemini 2.5 flash will be available to developers someday in June.
Finally, Google has introduced a model called Gemini Diffusion. This claims the company is “very fast.” It offers output 4-5 times faster than comparable models, comparable to the size of the model’s performance. Gemini Spreading is available from “Reliable Testers” today.