Google Rolls Out Gemini Deep Think Ai, A Reasoning Model That Tests Multiple Ideas In Parallel

Google DeepMind is rolling retired Gemini 2.5 Deep Think, which, nan institution says, is its astir precocious AI reasoning model, capable to reply questions by exploring and considering aggregate ideas simultaneously and past utilizing those outputs to take nan champion answer.

Subscribers to Google’s $250-per-month Ultra subscription will summation entree to Gemini 2.5 Deep Think successful nan Gemini app starting Friday.

First unveiled successful May astatine Google I/O 2025, Gemini 2.5 Deep Think is Google’s first publically disposable multi-agent model. These systems spawn AI aggregate agents to tackle a mobility successful parallel, a process that uses importantly much computational resources than a azygous agent, but tends to consequence successful amended answers.

Google utilized a variety of Gemini 2.5 Deep Think to score a golden medal astatine this year’s International Math Olympiad (IMO).

Alongside Gemini 2.5 Deep Think, nan institution says it is releasing nan exemplary it utilized astatine nan IMO to a prime group of mathematicians and academics. Google says this AI exemplary “takes hours to reason,” alternatively of seconds aliases minutes for illustration astir consumer-facing AI models. The institution hopes nan IMO exemplary will heighten investigation efforts, and intends to get feedback connected really to amended nan multi-agent strategy for world usage cases.

Google notes that nan Gemini 2.5 Deep Think exemplary is simply a important betterment complete what it announced astatine I/O. The institution besides claims to person developed “novel reinforcement learning techniques” to promote Gemini 2.5 Deep Think to make amended usage of its reasoning paths.

“Deep Think tin thief group tackle problems that require creativity, strategical readying and making improvements step-by-step,” said Google successful a blog station shared pinch TechCrunch.

The institution says Gemini 2.5 Deep Think achieves state-of-the-art capacity connected Humanity’s Last Exam (HLE) — a challenging trial measuring AI’s expertise to reply thousands of crowdsourced questions crossed math, humanities, and science. Google claims its exemplary scored 34.8% connected HLE (without tools), compared to xAI’s Grok 4, which scored 25.4%, and OpenAI’s o3, which scored 20.3%.

Google besides says Gemini 2.5 Deep Think outperforms AI models from OpenAI, xAI, and Anthropic connected LiveCodeBench6, a challenging trial of competitory coding tasks. Google’s exemplary scored 87.6%, whereas Grok 4 scored 79%, and OpenAI’s o3 scored 72%.

Benchmark scores. Image Credits: Google

Gemini 2.5 Deep Think automatically useful pinch devices specified arsenic codification execution and Google Search, and nan institution says it’s tin of producing “much longer responses” than accepted AI models.

In Google’s testing, nan exemplary produced much elaborate and aesthetically pleasing web improvement tasks compared to different AI models. The institution claims nan exemplary could assistance researchers and “potentially accelerate nan way to discovery.”

Art scenes made by Google’s AI (Credit: Google)

It seems that respective starring AI labs are converging astir nan multi-agent approach.

Elon Musk’s xAI precocious released a multi-agent strategy of its own, Grok 4 Heavy, which it says was capable to execute manufacture starring capacity connected respective benchmarks. OpenAI interrogator Noam Brown said connected a podcast that nan unreleased AI exemplary nan institution utilized to execute a golden badge astatine this year’s International Math Olympiad (IMO) was besides a multi-agent system. Meanwhile, Anthropic’s Research agent, which generates thorough investigation briefs, is besides powered by a multi-agent system.

Despite nan beardown performance, it seems that multi-agent systems are moreover costlier to service than accepted AI models. That intends tech companies whitethorn support these systems gated down their astir costly subscription plans, which xAI and now Google person chosen to do.

In nan coming weeks, Google says it plans to stock Gemini 2.5 Deep Think pinch a prime group of testers via nan Gemini API. The institution says it wants to amended understand really developers and enterprises whitethorn usage its multi-agent system.