As part of Google AI Ultra subscription, Gemini 2.5 Deep Think is a model designed for creative problem-solving through the use of parallel thinking techniques and extended inference time.
According to Google, Deep Think is designed to better solve problems requiring creativity, strategic planning, and a step-by-step approach. Those include iterative development and design, scientific and mathematical discovery, and algorithm development.
The currently available model is an improved and optimized variation of the Deep Think model that reached gold-medal performance at the 2025 International Mathematical Olympiad (IMO).
An advanced version of Gemini Deep Think solved five out of the six IMO problems perfectly, earning 35 total points, and achieving gold-medal level performance. The solutions can be found online here.
That variant could take hours to reason about complex problems, while the new one trades some reasoning depth for speed, making it more practical for day-to-day use, explains Google. Google is particularly keen to underscore the progress they made since the 2024 IMO competition, when their AlphaProof and AlphaGeometry models required experts to first translate problems from natural language into domain-specific languages and took three days to produce solutions.
Google explains that Gemini Deep Think generates multiple alternative solutions in parallel and revises or combines them throughout the reasoning process to arrive at the best answer. To make this approach feasible, the model requires an extended reasoning time window, which makes it unsuitable for interactive use, such as in a chat. As Google notes, the model may occasionally exhibit slowness or timeout issues.
Another key difference with other models in the Gemini family is the use of novel reinforcement learning techniques during training, which encourage the model to make use of these extended reasoning paths to test multiple hypotheses at once. Last but not least, the model also got access to a curated corpus of high-quality solutions to mathematics problems.
Google says Gemini 2.5 Deep Think achieves state-of-the-art performance on several benchmarks, including LiveCodeBench V6 and Humanity’s Last Exam.
One limitation early adopters pointed out on Reddit is the restricted number of queries available even to paying users, which started at five per day and was later doubled to ten. Some commenters take this as a sign of how expensive the model is to run, possibly equivalent to spinning up a large number of Gemini Pro models in parallel.
Like other models in its family, Gemini Deep Think follows the sparse mixture-of-experts (MoE) architecture and has native support for native multimodal support for text, vision, and audio inputs. It supports a 1M token input context window and 192K token output window.
In addition to the optimized Deep Think model available to AI Ultra subscribers, Google is also releasing the competition variant to the research community.
