Google Deepmind play Gemini 2.5 they thinkThe one, the company said, as the most advanced AI model, can answer the question by exploring and consider the various ideas to choose the best answers.
Customer to Google 250-monthly Ultra Subscriptions will get access to Gemini 2.5 in the following Gemini application starts on Friday.
Firstly open in May on Google I / O 2025, Gemini 2.5 The deepest is Google is the most common available multi-agent model. The system spread ai many agents to complete the questions in Parallels, the process that uses more resources than one agent, but tends to produce better answers.
Google uses a variety of gemini 2.5 deep Gold medal score In the international international year of international Mathematics this year (IMO).
Along with Gemini 2.5, the company said that releasing the model used in Imo to the mathematician group and academic. Google says this model AI is “takes several hours for a reason,” instead of seconds or minutes like a consumer’s ai model. The company hopes the IMO model will increase the research attempt, and the purpose is to get feedback on how to add multi-agent system for academic use case.
Google recorded a significant 2.5 gemini model of what is announced on i / o. This company also claimed to develop a novel technique “to encourage a very good use of the road.
“Deep thoughts can help people demand the problems that require creativity, strategic planning and make the steps improvements,” Google said in blog postcrch.
TechCrunch Events
San Francisco
|
October 27, 2025
The company said Gemini 2.5 Thoughts in the state of the state of the human end (HLE) – Trials that measure a Thouanese, Humanity, and Science. Google emphasizes the model print 34.8% in HLE (without tools), compared with GREK 4 Xai, which counts 25.4%, and opening 20.3%.
Google also said Gemini 2.5, Models are very deeply models from Openai, Xai, and anthqupik in livecodebench6, testing competitive tasks. Google Model print 87.6%, while the 4 grok scored 79%, and Opening O3 creates 72%.
Gemini 2.5 Automatically can be used automatically with a tool such as code execution, and the company that says it can produce “a longer traditional” response.
In Google Test, the model produces more detailed and aesthetic web development tasks as compared to other AI models. The company said the model could help the researchers and “chances of speeding the road to be found.”
It seems some of the major AI airways are conversion about multi-agents approach.
Xai Elon Muskububnd recently released its own multi-agent agent system, The grilled 4 heavyThat said that can achieve the main performance in some benchmarks. Opening Noam Research says podcast If this ai model is not considered this company to obtain a gold medal in the international international Olympics this year also includes multi Agent systems. Meanwhile, Anticroprapic Research AgentThat generates a study summary, also supported by multi agent systems.
Although strong performance, it seems that multi-agent agents are even more expensive than traditional AI AI models. That means tech companies can keep the system for the most expensive subscription plan, the Xai and now Google selects.
The coming weeks, Google said a plan to show Gemini 2.5, think deeply with thesis tests through Gemini API. The company wants to know how the developer and the company can use the multi-agent system.