Google rolls out Gemini Deep Think AI, a reasoning model that tests multiple ideas in parallel

Google Dipmind Roll Out Gemini 2.5 deep thoughtsThe company that says, this is the most advanced AI argument model, able to answer and consider multiple ideas simultaneously, and then use those results to choose the best answer.

Google $ 250 per month customers Ultra Subscription will gain access to Jeminely 2.5 deep thoughts in the Jemi application from Friday.

Google I/& O 2025 first unveiled in May, Gemi 2.5 Deep Think is Google’s first publicly available multi-agent model. These systems are parallel to a question in parallel, AI span multiple agents, a process that uses significantly more calculating resources than a single agent, but as a result of better answer.

Google Jemini has used a variation of 2.5 dip thank Score a gold medal This year’s International Mathematics Olympiad (IMO).

In addition to Jemi 2.5 Deep Think, the company has said that the model used in the IMO is publishing a selected team of mathematicians and academics. Google says this AI model is “hours to argue” instead of a few seconds or minutes like most customer-face AI models. The company hopes that the IMO model will enhance research efforts and notice how to improve multiple-agent systems in academic use.

Google notes that Gemi 2.5 Deep Think Model is a significant improvement than what it announced in I/O. The company also claimed that the company developed “the novel reinforcement learning techniques” to encourage Gemi 2.5 deeply to think about better use of its rational paths.

In a blog post shared with TechCrunch, Google said, “Deep Think can help people deal with problems for which creativity, strategic plans and improvements have stepped in step by step.

The company says that Gemi 2.5 has achieved sophisticated performance on the last examination (HLE) of the Deep Think Humanity-a challenging test of measuring AI’s skills to answer thousands of people across science and science. Google claims that its model has scored 34.8%in HLE (without equipment), compared to Jai Grock 4, which is 25.4%and OpenAE, which scored 20.3%.

Google also says that Gemi 2.5 Deep Thinks Open AAE, Jai and anthropological AI models out of Livcodbench 6, a challenging test of competitive coding functions. Google’s model scored 87.6%, where Grock scored 49%and scored OpenAEI and 3 72%.

Gemi 2.5 Deep Think works with equipment such as code execution and Google search and the agency says it is capable of producing “much longer response” compared to the Traditional AI model.

In the Google exam, the model has developed the work of developing a more detailed and aesthetic web than other AI models. The company has claimed that the model can help researchers and “possiblely accelerate the path of invention.”

It looks like a number of top AI labs are converting around the multi-agent approach.

Elon Mask’s Jai recently released a multi-agent system, Grock is 4 heavyWhich has been said that several criteria were able to achieve the top performance of the industry. OpenAI researcher Noam Brown says a Podcast The company was also a multi-agent system to achieve the gold medal at the International Mathematics Olympiad (IMO) this year. In the me, Ethnographic research agentWhich produces a brief summary of the study, it is powered by a multi-agent system.

Despite the strong performance, it seems that multi-agent systems are even expensive to serve than the traditional tidal AI model. Its money is the technology companies these systems can keep them behind the most expensive subscription plans, which have chosen to do Google and now Google.

In the coming weeks, Google has said that it is planned to share with a selected examiners of Gemi 2.5 Deep Thinks via Gemi API. The company says that it is better to understand how the developer and initiatives can use its multi-agent systems.