Google DeepMind is rolling out Gemini 2.5 Deep Think, which the company describes as its most advanced AI reasoning model. This system can answer questions by exploring multiple ideas simultaneously and selecting the best response. Subscribers to Google’s $250-per-month Ultra subscription will gain access to Gemini 2.5 Deep Think in the Gemini app starting Friday.
First unveiled in May at Google I/O 2025, Gemini 2.5 Deep Think is Google’s first publicly available multi-agent model. These systems deploy multiple AI agents to tackle questions in parallel, using more computational resources but often yielding better results. Google used a specialized version of this model to win a gold medal at this year’s International Math Olympiad (IMO).
Alongside the public release, Google is providing the IMO-specific model to a select group of mathematicians and academics. Unlike consumer-facing AI models that respond in seconds, this version takes hours to reason. Google hopes it will advance research and gather feedback to refine multi-agent systems for academic use.
Google claims Gemini 2.5 Deep Think is a major upgrade over its initial I/O announcement, with new reinforcement learning techniques to improve reasoning efficiency. The company states that the model excels in creativity, strategic planning, and step-by-step problem-solving.
In benchmarks, Gemini 2.5 Deep Think reportedly outperforms competitors. It achieved 34.8% on Humanity’s Last Exam (HLE), surpassing xAI’s Grok 4 (25.4%) and OpenAI’s o3 (20.3%). It also scored 87.6% on LiveCodeBench6, a competitive coding test, compared to Grok 4’s 79% and OpenAI’s o3’s 72%.
The model integrates tools like code execution and Google Search and can generate longer, more detailed responses. Google’s tests suggest it produces higher-quality web development outputs and could accelerate research breakthroughs.
Other AI labs are also embracing multi-agent systems. xAI recently launched Grok 4 Heavy, while OpenAI and Anthropic have developed similar architectures. However, these systems are expensive to run, leading companies to restrict access to premium subscriptions.
Google plans to share Gemini 2.5 Deep Think with select testers via its API in the coming weeks, aiming to explore potential applications for developers and enterprises.