Google rolls out Gemini Deep Think AI, a reasoning model that tests multiple ideas in parallel

Google DeepMind has launched its latest AI innovation, Gemini 2.5 Deep Think, touted as the company's most sophisticated reasoning model to date. This groundbreaking system is capable of evaluating multiple concepts simultaneously to provide optimal answers. Starting this Friday, users subscribed to Google's $250-per-month Ultra plan will have exclusive access to Gemini 2.5 Deep Think through the Gemini app. First introduced during the Google I/O event in May 2025, Gemini 2.5 Deep Think marks a significant milestone as Google's inaugural publicly available multi-agent model. This advanced model employs multiple AI agents to collaboratively address inquiries, a method that demands considerable computational power but often yields superior results. Notably, a variant of this model helped secure a gold medal at the recent International Math Olympiad (IMO). In addition to Gemini 2.5 Deep Think, Google is also releasing the specific model used at the IMO to a limited audience of mathematicians and researchers. According to the tech giant, this AI model requires several hours to complete its reasoning, a stark contrast to the rapid response times typical of consumer-focused AI solutions. Google anticipates that this academic version will bolster research initiatives and seeks feedback to refine the multi-agent framework for scholarly applications. The enhancements in Gemini 2.5 Deep Think compared to its predecessor are noteworthy. Google claims to have implemented innovative reinforcement learning techniques that enable the model to optimize its reasoning pathways more effectively. "Deep Think is designed to assist in addressing challenges that demand creativity, strategic foresight, and iterative improvements," Google articulated in a blog post to TechCrunch. Performance metrics reveal that Gemini 2.5 Deep Think has achieved state-of-the-art results on Humanity’s Last Exam (HLE), a rigorous assessment measuring AI proficiency across diverse fields including mathematics and humanities. The model reportedly scored 34.8% on the HLE, significantly surpassing competitors, such as xAI’s Grok 4 at 25.4% and OpenAI’s o3 at 20.3%. On the LiveCodeBench6 coding challenge, Gemini 2.5 Deep Think excelled with a score of 87.6%, outperforming Grok 4 and o3, which received scores of 79% and 72%, respectively. Moreover, Gemini 2.5 Deep Think seamlessly integrates with various tools, including code execution capabilities and Google Search, and is capable of generating more extensive and detailed outputs than many existing AI models. Google's evaluations indicate that the model excels in producing aesthetically pleasing web development outcomes compared to its rivals, potentially paving the way for accelerated discoveries in research fields. The trend of multi-agent systems is gaining traction among leading AI laboratories. xAI, founded by Elon Musk, has recently introduced its own multi-agent system, Grok 4 Heavy, claiming to achieve high performance on multiple benchmarks. Additionally, OpenAI has hinted at utilizing a multi-agent model for its gold medal-winning performance at the IMO, and Anthropic is employing a similar system for its Research agent that produces comprehensive research briefs. Despite their impressive capabilities, the operational costs associated with multi-agent systems are notably higher than traditional AI models. As a result, companies may restrict access to these advanced systems to their premium subscription tiers, as evidenced by xAI and now Google. In the upcoming weeks, Google plans to extend access to Gemini 2.5 Deep Think to a select group of testers via the Gemini API, aiming to explore potential applications in development and enterprise settings.