Google Deepmind exceeds Gemini 2.5 Deep Think, which, he says, is AI’s most advanced logic model, capable of answering questions while examining and examining multiple ideas and then using these outputs to choose the best answer.
Ultra subscription subscribers of $ 250 per month will gain access to the Gemini 2.5 Deep Think to the Gemini application that begins on Friday.
First revealed in May to Google I/O 2025, Gemini 2.5 Deep Think is Google’s first publicly available multi -factor model. These systems create multiple AI agents to address a question at the same time, a process that uses significantly more computational resources than a single agent, but tends to lead to better answers.
Google used a variant of the Gemini 2.5 Deep Think scoring a gold medal at this year’s International Mathematics Olympiad (IMO).
Along with the Gemini 2.5 Deep Think, the company says it releases the model it used in IMO in a selected group of mathematicians and academics. Google says this AI model “lasts for logic” instead of seconds or minutes, like most AI models facing consumers. The company hopes that the IMO model will enhance research efforts and aims to receive feedback on how to improve the multiple factors system for cases of academic use.
Google notes that the Gemini 2.5 Deep Think model is a significant improvement compared to what it announced in I/O. The company also claims to have developed “new aid learning techniques” to encourage Gemini 2.5 Deep Think to make better use of its logic paths.
“Deep thinking can help people deal with problems that require creativity, strategic planning and step -by -step improvements,” Google told a blog shared with TechCrunch.
TechCrunch event
Francisco
|
27-29 October 2025
The company says that the Gemini 2.5 Deep Think achieves the state-of-the-art performance in the latest HEL (HLE) exams-a provocative test that measured AI’s ability to answer thousands of questions in mathematics, humanities and sciences. Google claims that its model scored 34.8%in HLE (without tools) compared to Xai’s Grok 4, which recorded 25.4%and Openai’s O3, which recorded 20.3%.
Google also says that the Gemini 2.5 Deep Think surpasses AI models from Openai, XAI and Anthropic in Livecodebench 6, a difficult trial of competing coding. Google’s model was 87.6%, while Grok 4 scored 79%and Openai’s O3 recorded 72%.
Gemini 2.5 Deep thinks automatically works automatically with tools such as code execution and Google search and the company says it is able to produce “much greater answers” than traditional AI models.
In the Google test, the model produced more detailed and aesthetically pleasing tissue development tasks compared to other AI models. The company claims that the model could help researchers and “may accelerate the course to discovery”.
It seems that many top AI laboratories converge around the multiple factors approach.
Elon Musk’s XAI has recently released its own multi -factor system, the Grok 4 Heavy, which says it was able to achieve the top performance at various points of reference. Openai Noam Brown researcher told a podcast That the non -circulated AI model used by the company to achieve a gold metal at this year’s International Olympia Mathematics was also a multi -factor system. Meantime, Anthropic’s research agentwhich creates detailed surveys, is also powered by a multi -factor system.
Despite the strong performance, it seems that multi -factor systems are even more expensive to serve traditional AI models. This means that technology companies can keep these systems behind their most expensive subscription plans that Xai and now Google has chosen to do.
In the coming weeks, Google says it is planning to share the Gemini 2.5 Deep Think with a selected tester team through API Gemini. The company says it wants to better understand how developers and businesses can use the multi -factor system.
