Gemini 2.5 Deep Think

What it is

Gemini 2.5 Deep Think is an enhanced reasoning mode for Gemini 2.5 Pro. Where an ordinary model produces an answer in a single pass, this mode considers multiple hypotheses before responding^[1]. Google announced it as an experimental feature at Google I/O on 20 May 2025^[1].

How it works

The mode is built around parallel thinking. Instead of following one chain of reasoning from start to finish, it generates many ideas at once and then revises or combines them into a final answer^[2]. Google also encouraged longer inference time through reinforcement learning, so that the model keeps working on a problem rather than stopping at the first plausible answer^[2].

Most language models spend roughly the same amount of computation on every question they are asked. A reasoning mode changes that balance: more computation is spent after the question is posed and before the answer is given, so that harder problems get more room to be worked through. That is what the extended inference time is for, and it is why a reasoning mode can behave differently from the same underlying model answering directly.

Announcement and first release

At I/O, Google reported that the mode led LiveCodeBench and scored 84.0% on MMMU^[1]. It was not shipped to everyone at once: the first release went to trusted testers through the Gemini API, so that safety evaluation could happen before wider availability^[1].

On 1 August 2025 the mode arrived in the Gemini app for Google AI Ultra subscribers, with a fixed number of prompts per day^[2].

The olympiad result

In July 2025 an advanced version of Gemini with Deep Think solved five of the six problems at the International Mathematical Olympiad^[3]. Google DeepMind said the model solved those five perfectly, for 35 total points, and that the results were officially graded and certified by IMO coordinators using the same criteria applied to student solutions^[3]. Google DeepMind dated its announcement post 21 July 2025^[3].

The version that reached the app is a faster variation of the model behind that result. Google described it as a faster variation of a model that takes hours to reason and that reached gold-medal standard at the 2025 International Mathematical Olympiad, adding that in internal evaluations the app version reached bronze level on the IMO 2025 benchmark^[2]. The official gold-medal model itself was shared with a small group of mathematicians and academics^[2].

Where it sits

Deep Think is best understood as a mode rather than a separate model line: it belongs to Gemini 2.5 Pro, with extra reasoning computation placed before the answer and with its own rules of access — trusted testers and the API first, then paying subscribers with a daily prompt limit^[1][2]. The olympiad result matters because the grading was done by IMO coordinators under the same criteria used for students, which makes the comparison to a certified medal standard a concrete one rather than a marketing claim^[3]. For users of the Gemini app, the trade-off is visible in the design: a slower, more deliberate way of answering is offered as an option with a fixed daily budget of prompts^[2].