Google DeepMind is rolling out Gemini 2.5 Deep Think, which, the corporate says, is its most superior AI reasoning mannequin, in a position to reply questions by exploring and contemplating a number of concepts concurrently after which utilizing these outputs to decide on one of the best reply.

Subscribers to Google’s $250-per-month Ultra subscription will acquire entry to Gemini 2.5 Deep Assume within the Gemini app beginning Friday.

First unveiled in Might at Google I/O 2025, Gemini 2.5 Deep Assume is Google’s first publicly obtainable multi-agent mannequin. These methods spawn AI a number of brokers to deal with a query in parallel, a course of that makes use of considerably extra computational assets than a single agent, however tends to end in higher solutions.

Google used a variation of Gemini 2.5 Deep Assume to score a gold medal at this yr’s Worldwide Math Olympiad (IMO).

Alongside Gemini 2.5 Deep Assume, the corporate says it’s releasing the mannequin it used on the IMO to a choose group of mathematicians and teachers. Google says this AI mannequin “takes hours to motive,” as an alternative of seconds or minutes like most consumer-facing AI fashions. The corporate hopes the IMO mannequin will improve analysis efforts, and goals to get suggestions on tips on how to enhance the multi-agent system for tutorial use instances.

Google notes that the Gemini 2.5 Deep Assume mannequin is a major enchancment over what it introduced at I/O. The corporate additionally claims to have developed “novel reinforcement studying methods” to encourage Gemini 2.5 Deep Assume to make higher use of its reasoning paths.

“Deep Assume will help individuals deal with issues that require creativity, strategic planning and making enhancements step-by-step,” mentioned Google in a weblog publish shared with TechCrunch.

The corporate says Gemini 2.5 Deep Assume achieves state-of-the-art efficiency on Humanity’s Final Examination (HLE) — a difficult take a look at measuring AI’s means to reply hundreds of crowdsourced questions throughout math, humanities, and science. Google claims its mannequin scored 34.8% on HLE (with out instruments), in comparison with xAI’s Grok 4, which scored 25.4%, and OpenAI’s o3, which scored 20.3%.

Google additionally says Gemini 2.5 Deep Assume outperforms AI fashions from OpenAI, xAI, and Anthropic on LiveCodeBench6, a difficult take a look at of aggressive coding duties. Google’s mannequin scored 87.6%, whereas Grok 4 scored 79%, and OpenAI’s o3 scored 72%.

Gemini 2.5 Deep Assume robotically works with instruments equivalent to code execution and Google Search, and the corporate says it’s able to producing “for much longer responses” than conventional AI fashions.

In Google’s testing, the mannequin produced extra detailed and aesthetically pleasing internet growth duties in comparison with different AI fashions. The corporate claims the mannequin may support researchers and “doubtlessly speed up the trail to discovery.”

Evidently a number of main AI labs are converging across the multi-agent method.

Elon Musk’s xAI not too long ago launched a multi-agent system of its personal, Grok 4 Heavy, which it says was in a position to obtain trade main efficiency on a number of benchmarks. OpenAI researcher Noam Brown mentioned on a podcast that the unreleased AI mannequin the corporate used to realize a gold medal at this yr’s Worldwide Math Olympiad (IMO) was additionally a multi-agent system. In the meantime, Anthropic’s Research agent, which generates thorough analysis briefs, can be powered by a multi-agent system.

Regardless of the sturdy efficiency, it appears that evidently multi-agent methods are even costlier to serve than conventional AI fashions. Meaning tech corporations might hold these methods gated behind their most costly subscription plans, which xAI and now Google have chosen to do.

Within the coming weeks, Google says it plans to share Gemini 2.5 Deep Assume with a choose group of testers by way of the Gemini API. The corporate says it needs to higher perceive how builders and enterprises might use its multi-agent system.