Google DeepMind released Gemini 2.5 Deep Think on August 1, 2025, the company's most advanced AI reasoning system capable of generating and evaluating multiple ideas simultaneously before selecting the best solution. This multi-agent system, first unveiled at Google I/O in May 2025, uses significantly more computational resources than traditional AI models and is now available to Google's $250-per-month Ultra subscribers.
Gemini 2.5 Deep Think delivers exceptional performance on key benchmarks, scoring 34.8% on Humanity's Last Exam (HLE) without tools, compared to xAI's Grok 4 at 25.4% and OpenAI's o3 at 20.3%, as well as 87.6% on LiveCodeBench 6 coding test, outperforming Grok 4's 79% and OpenAI's o3's 72%. Google also developed an enhanced version that achieved a gold medal standard at this year's International Mathematical Olympiad (IMO), and this variation, which takes hours to process complex problems, is being shared with a select group of mathematicians and academics for further research purposes. The company implemented novel reinforcement learning techniques to encourage the model to make better use of its reasoning paths.
Gemini Deep Think automatically works with tools such as code execution and Google Search, and according to the company, can produce much longer responses than traditional AI models. Google plans to share Gemini 2.5 Deep Think with select testers via the Gemini API in the coming weeks to better understand how developers and enterprises might use this multi-agent system. This development reflects a broader industry trend, as several leading AI labs – including xAI and OpenAI – are converging around the multi-agent approach, though due to high computational costs, these systems are likely to remain behind the most expensive subscription tiers.
Sources:
Maxwell Zeff
The Deep Think team
Andre Revilla
