Google Launches Gemini AI for Advanced Parallel Reasoning
▼ Summary
– Google DeepMind launched Gemini 2.5 Deep Think, its most advanced AI reasoning model, capable of exploring multiple ideas simultaneously to choose the best answer.
– The model will be available to subscribers of Google’s $250-per-month Ultra plan starting Friday, marking its first public release as a multi-agent system.
– Google used a variant of Gemini 2.5 Deep Think to win a gold medal at the International Math Olympiad (IMO) and is releasing this specialized model to select academics for research feedback.
– Gemini 2.5 Deep Think outperforms competitors like OpenAI and xAI on benchmarks such as Humanity’s Last Exam (HLE) and LiveCodeBench6, achieving higher accuracy in complex tasks.
– Multi-agent AI systems, though more resource-intensive, are gaining traction among leading labs like Google, xAI, and OpenAI, but their high costs may limit access to premium subscriptions.
Google has unveiled Gemini 2.5 Deep Think, its most sophisticated AI reasoning model yet, designed to tackle complex problems by evaluating multiple solutions simultaneously. This advanced system, now available to subscribers of Google’s $250-per-month Ultra plan, marks a significant leap in AI capabilities, particularly in fields requiring deep analytical thinking.
First introduced at Google I/O 2025, Gemini 2.5 Deep Think represents the company’s first public multi-agent AI model, where multiple AI agents work in parallel to solve problems. While this approach demands greater computational power, it delivers superior results compared to single-agent systems. Google demonstrated its potential by leveraging a specialized version of the model to secure a gold medal at this year’s International Math Olympiad (IMO).
Alongside the consumer release, Google is providing access to its IMO-grade model to a select group of mathematicians and researchers. Unlike standard AI models that generate responses in seconds, this version requires hours of reasoning, a deliberate design choice aimed at advancing academic research. The company hopes feedback from experts will refine the system for specialized applications.
Google emphasizes that Gemini 2.5 Deep Think outperforms earlier iterations announced at I/O, thanks to new reinforcement learning techniques that enhance its reasoning efficiency. According to the company, the model excels in tasks requiring creativity, strategic planning, and iterative improvements, making it a valuable tool for both professionals and researchers.
In benchmark tests, the model achieved state-of-the-art results on Humanity’s Last Exam (HLE), scoring 34.8% without external tools, outperforming competitors like xAI’s Grok 4 (25.4%) and OpenAI’s o3 (20.3%). It also dominated LiveCodeBench6, a competitive coding assessment, with an 87.6% success rate compared to Grok 4’s 79% and OpenAI’s 72%.
Beyond raw performance, Gemini 2.5 Deep Think integrates seamlessly with tools like code execution and Google Search, enabling longer, more detailed responses. Early tests suggest it produces higher-quality outputs for web development tasks, potentially accelerating innovation in research and software engineering.
The multi-agent approach is gaining traction among AI leaders. xAI recently launched Grok 4 Heavy, while OpenAI and Anthropic have also adopted similar architectures for high-performance applications. However, the computational costs of these systems remain steep, likely restricting access to premium-tier subscriptions, a trend already seen with xAI and now Google.
In the coming weeks, Google plans to offer Gemini 2.5 Deep Think to select developers via its API, aiming to explore real-world applications in enterprise and research settings. As AI continues to evolve, multi-agent systems like this could redefine how complex problems are approached across industries.
(Source: TechCrunch)
