Today, Google announced the rollout of the Gemini 2.5 Deep Think model in the Gemini app for Google AI Ultra subscribers. This new model beats OpenAI"s o3 and xAI"s Gork 4 models in several key benchmarks.
Back in May, at Google I/O 2025, Google first announced the Gemini 2.5 Pro Deep Think mode, which used new research techniques to consider multiple hypotheses before responding. Today"s rollout is an updated Deep Think model that includes improvements based on feedback from early trusted testers and recent research breakthroughs. Google claims that this updated Deep Think model is a significant improvement over the version first announced at I/O.
Google also revealed that this new Deep Think model is a variation of the model that recently achieved the gold-medal standard at this year’s International Mathematical Olympiad (IMO). To make it suitable for everyday usage, Google has made it faster, and it, therefore, reaches only Bronze-level performance on the 2025 IMO benchmark.
As seen in the benchmarks in the screenshot above, Gemini 2.5 Deep Think is shown to achieve excellent performance across LiveCodeBench V6, Humanity’s Last Exam, IMO 2025, and AIME 2025.
Google AI Ultra subscribers can now use Deep Think in the Gemini app with a fixed number of prompts per day by using the "Deep Think" option that appears in the prompt bar when the Gemini 2.5 Pro model is selected. Google noted that Deep Think automatically works with tools such as code execution and Google Search. Google is also planning to release Deep Think with and without tools to a set of trusted developers via the Gemini API in the coming weeks.
The big question now is whether the upcoming OpenAI GPT-5 model will outperform Gemini 2.5 Deep Think in key benchmark results.