Google Rolls Out Gemini Think Ai, a reasoning model that tries many ideas in parallel

Google Deepmind is rolled Gemini 2.5 Think deepthat, the company says, it is his most advanced about answering the questions of ideas simultaneously and then using those best outputs to choose the best answer.

Subscribers to $ 250-per-month Ultra The subscription has turned on Gemini 2.5 Gemini access in the Gemini app that begins on Friday.

First Sent in May in Google I / O 2025, Gemini 2.5 Think is the first Google model publicly publicly. These spawn systems are agents to treat a request in parallel, a process using significant resources significantly than another agent, but tends as best answers.

Google used a variation of Gemini 2.5 Think deep Score a gold medal In this international math of international math (imo).

To Corsormid Gemini 2.5 Profands think, the company says it’s ReSUPDDY the Model, used to a group of mathematical and accommodation. Google says ai model “takes hours to reason,” instead of seconds or minutes as most companies ai. The company is hoping that the imo model will strengthen the research efforts, and the purpose of obtaining feedback on how to improve the multi-agent system for academic use cases.

Google notes that the Gemini 2.5 the pattern to think a significant improvement of what you have announced to I / o. The company also tells you “to encourage 2.5 profiles to do better use their railing paths.

“Think deep can help people add creativity, strategized planning and make the prettier passed parties.” He said: Google in a sharing post shared with techcrunnch.

The company tells Gemini 2.5 depths of the State of the State of the State of the State of the State (HLE) – a test of humanity that measures the ability to the hairs in the math, and science. Google declares her model scored 34.8% on HLE (without instruments), compare to Grook Xai), which has done 25.4%, and Oppairs of 20.3%.

Google Des Gemini 2.5 Think Proford the Ai, Saint Antropica on Livecadowbench6, a challenging test of compulsory codation. The Google model scored 87.6%, while Grook 4 has scored 79%, and at Openai’s O3 has scored 72%.

Gemini 2.5 they think the utterance tools as the Google code and research, and the company is capable of producing production models “than tradition.

In the Google test, the model produced the most detailed development and aesthetically pleasing development The company declares the pattern could help researchers and “accelerate potentially the walk to break up.”

It looks like many point labs are convergent around the multi-agent approach.

Xai of Elon Musk recently published a multi-agent system of their own, Grook 4 heavywho says he was able to get performance leading to many benchmarks. Rannober Openai Noam Brown told about a podcast May the model useless the company used to obtain a gold medal in this year’s international Olympics (imo) was also a multi-agent system. In your me? Anthropic search agentthat generates short research, is also fueled by a multi-agent system.

Despite the strong spectator, it seems that multi-aent systems systems is also costlier to serve the models of the traditional AI traditional. That means technical companies can keep these systems gratified behind their most expensive underline, which xai and now Google has selected.

Nesties are Google, says they were thinking about sharing Gemini 2.5 I am thinking of a selected group of wishes via API topics. The company says he wants to understand better as developers and business can use their multi-agent system.