Google / Elyse Betters Picaro / ZDNET
Two weeks ago, Google and OpenAI touted their models’ award-winning performance at the International Math Olympiad (IMO). Now, Google is making a version of its model available to the public.
Also: This one feature could make GPT-5 a true game changer (if OpenAI gets it right)
On Friday, Google launched Deep Think in the Gemini app for Google Ultra subscribers, a premium subscription tier that costs $250 per year or $125 for the first three months. Although the model is a variation of the one that achieved the gold-medal standard at IMO, it is faster for everyday tasks. Internal evaluations suggested the model reaches bronze-level on the 2025 IMO benchmark.
How does it work
The superior performance solving complex problems is enabled by a parallel thinking technique, which allows the model to simultaneously generate and process multiple ideas, even combining different ones as necessary to find the best answer.
Also: Google upgrades AI Mode with Canvas and 3 other new features – how to try them
Other factors contributing to the high performance include an extended inference time, also known as thinking time, which allows Deep Think to explore more options before arriving at an answer, and new reinforcement learning techniques that help the model to become a better problem-solver over time.
According to Google, Deep Think excels at iterative development and design, as seen in the image above, scientific and mathematical discovery, and coding. These results are reflected in Gemini 2.5 Deep Think’s performance across state-of-the-art benchmarks, including Humanity’s Last Exam, an exam with multi-modal questions in over 100 subjects, such as math, science, and the humanities.
Google also shared that the Gemini 2.5 Deep Think has shown better content safety and tone-objectivity compared to Gemini 2.5 Pro, with the caveat that it denied benign requests at a higher rate.
How to access
Google AI Ultra subscribers can access Deep Think in the Gemini app with a fixed set of prompts daily. To select the model, toggle “Deep Think” in the prompt bar when selecting 2.5 Pro on the model selector. The company also shared that it’s working to release Deep Think, with and without tools, to a set of trusted testers via the Gemini API in the coming weeks.
Also: OpenAI teases imminent GPT-5 launch. Here’s what to expect
The Gemini 2.5 Deep Think model that achieved the gold-medal standard will be shared with a small group of mathematicians and academics. The intention is that this model will be used to advance their work, and it is hoped the lessons will provide feedback for improvements.
Want more stories about AI? Check out AI Leaderboard, our weekly newsletter.
