Google Releases Its Award-winning Math Olympiad Model, But It'll Come At A Price

Google

Two weeks ago, Google and OpenAI touted their models' award-winning capacity astatine nan International Math Olympiad (IMO). Now, Google is making a type of its exemplary disposable to nan public. 

Also: This 1 characteristic could make GPT-5 a existent crippled changer (if OpenAI gets it right)

On Friday, Google launched Deep Think successful nan Gemini app for Google Ultra subscribers, a premium subscription tier that costs $250 per twelvemonth aliases $125 for nan first 3 months. Although nan exemplary is simply a variety of nan 1 that achieved nan gold-medal modular astatine IMO, it is faster for mundane tasks. Internal evaluations suggested nan exemplary reaches bronze-level connected nan 2025 IMO benchmark.

How does it work

The superior capacity solving analyzable problems is enabled by a parallel reasoning technique, which allows nan exemplary to simultaneously make and process aggregate ideas, moreover combining different ones arsenic basal to find nan champion answer. 

Also: Google upgrades AI Mode pinch Canvas and 3 different caller features - really to effort them

Other factors contributing to nan precocious capacity see an extended conclusion time, besides known arsenic reasoning time, which allows Deep Think to research much options earlier arriving astatine an answer, and caller reinforcement learning techniques that thief nan exemplary to go a amended problem-solver complete time.

Google

According to Google, Deep Think excels astatine iterative improvement and design, arsenic seen successful nan image above, technological and mathematical discovery, and coding. These results are reflected successful Gemini 2.5 Deep Think's capacity crossed state-of-the-art benchmarks, including Humanity's Last Exam, an exam pinch multi-modal questions successful complete 100 subjects, specified arsenic math, science, and nan humanities.

Google

Google besides shared that nan Gemini 2.5 Deep Think has shown amended contented information and tone-objectivity compared to Gemini 2.5 Pro, pinch nan caveat that it denied benign requests astatine a higher rate. 

How to access

Google AI Ultra subscribers tin entree Deep Think successful nan Gemini app pinch a fixed group of prompts daily. To prime nan model, toggle "Deep Think" successful nan punctual barroom erstwhile selecting 2.5 Pro connected nan exemplary selector. The institution besides shared that it's moving to merchandise Deep Think, pinch and without tools, to a group of trusted testers via nan Gemini API successful nan coming weeks. 

Also: OpenAI teases imminent GPT-5 launch. Here's what to expect

The Gemini 2.5 Deep Think exemplary that achieved nan gold-medal modular will beryllium shared pinch a mini group of mathematicians and academics. The volition is that this exemplary will beryllium utilized to beforehand their work, and it is hoped nan lessons will supply feedback for improvements.