Google DeepMind Unveils Enhanced Gemini 2.5 Deep Think: The Next Step in Multi-Agent AI Technology

Google DeepMind has unveiled Gemini 2.5 Deep Think, which the company claims to be the most advanced artificial intelligence model capable of addressing inquiries by exploring and examining multiple concepts simultaneously, subsequently utilizing the findings to determine the best response.

Starting Friday, subscribers to Google Ultra will gain access to Gemini 2.5 Deep Think via the Gemini application.

Initially showcased in May at the Google I/O 2025 conference, Gemini 2.5 Deep Think represents Google’s first publicly available multi-agent model. These systems employ several AI agents to tackle problems in parallel. This approach requires significantly more computational power compared to a single-agent operation but generally yields higher quality results.

Google utilized a variant of Gemini 2.5 Deep Think to secure a gold medal at this year’s International Mathematical Olympiad (IMO).

The company states that alongside Gemini 2.5 Deep Think, they are providing access to the model employed at the IMO to a select group of mathematicians and researchers. Google hopes the IMO model will facilitate further research and aims to gather feedback on improving the multi-agent system for academic purposes.

Gemini 2.5 Deep Think has undergone significant enhancements since its initial announcement at I/O. The company also claims to have developed «new reinforcement learning methods» that enable Gemini 2.5 Deep Think to optimize its reasoning pathways more effectively.

«Deep Think can assist individuals in tackling problems that require creative solutions, strategic planning, and iterative improvement,» Google stated in a blog post published on TechCrunch.

The company asserts that Gemini 2.5 Deep Think achieves outstanding performance on the «Humanity’s Last Exam» (HLE), a challenging assessment designed to evaluate an AI’s ability to answer thousands of questions across mathematics, the humanities, and the natural sciences sourced from a crowdsourcing platform. The model scored 34.8% on the HLE (without using tools), while xAI’s Grok 4 scored 25.4%, and OpenAI’s G3 achieved 20.3%.

Google further claims that Gemini 2.5 Deep Think outperforms AI models from OpenAI, xAI, and Anthropic in LiveCodeBench6, a demanding programming tasks test. The Google model obtained 87.6%, compared to Grok 4’s 79% and OpenAI’s o3’s 72%.

Gemini 2.5 Deep Think automatically integrates with tools like code execution and Google Search. The company asserts that this system can generate «far more elaborate responses» than traditional AI models.

During testing, Google’s model executed more detailed and visually appealing web development tasks in comparison to other AI models. The company claims this model could assist researchers and «potentially expedite discovery processes.»

According to Google, in the upcoming weeks, the company plans to grant access to Gemini 2.5 Deep Think to a select group of testers through the Gemini API. The goal is to better understand how developers and businesses can leverage its multi-agent system.

Delegate some of your routine tasks with BotHub! No VPN is required to access the service, and Russian cards are accepted. You can get 100,000 free tokens for your initial tasks and start working with neural networks right away!

For the original source of this news, click here.