Artificial Intelligence upgrade, Gemini 2.5 Deep Think, set to be accessible for Google AI Ultra consumers
==========================================================================
Google's latest offering, Gemini 2.5 Deep Think, is making waves in the world of artificial intelligence. This advanced reasoning model is now being rolled out to more users, providing them with the tools to tackle complex problems.
Before its wide release, Deep Think was fine-tuned by a small group of trusted testers through the Gemini API. The model was built on the latest cutting-edge research and can weigh multiple ideas before landing on a response, much like OpenAI's top models.
In internal IMO tests, Deep Think scored a Bronze level, indicating its performance on everyday tasks. After passing safety checks and receiving expert input, it went on to score big on the 2025 U.S. Math Olympiad and Humanity's Last Exam, a challenging test covering advanced science and math. It even pulled an impressive 84% on the MMMU test and crushed Olympiad questions, although some numbers were kept quiet.
One of the key differences between Deep Think and OpenAI's models lies in its multi-agent "parallel thinking" architecture. This enables Deep Think to explore multiple ideas simultaneously rather than sequentially, resulting in improved reasoning depth and the ability to solve complex problems by considering and refining multiple solutions in parallel.
Functionally, Deep Think is natively multimodal, handling text, images, audio, and video inputs within an extremely large context window (up to 1 million tokens), and it can generate very long text outputs (up to 192,000 tokens). This allows it to manage tasks requiring extensive context and long reasoning chains.
In terms of performance, Deep Think achieves state-of-the-art results on challenging benchmarks such as LiveCodeBench V6 (coding), Humanity’s Last Exam, and it was instrumental in earning a gold medal for Google at the International Math Olympiad 2025. The model uses novel reinforcement learning techniques to better leverage its extended reasoning paths, improving over time with use.
The Google AI Ultra plan, which includes access to Gemini 2.5 Deep Think, offers more features and benefits compared to the Google AI Pro plan. The Ultra plan costs £250 per month and includes early access to experimental video models, higher limits in tools like NotebookLM, an AI agent manager prototype, and 30TB of storage.
In conclusion, Gemini 2.5 Deep Think outperforms OpenAI’s models on complex reasoning, multimodal input handling, long-context reasoning, and creative problem solving, enabled by its multi-agent parallel thinking and novel training methods, while OpenAI models typically follow more conventional single-agent, sequential reasoning architectures.
[1] Google Research Blog: Introducing Gemini 2.5 DeepMind [2] TechCrunch: Google's new AI model, DeepMind, is a big deal for the industry [3] VentureBeat: Google's Gemini 2.5 DeepMind is a major step forward for AI research [4] ArXiv: Gemini 2.5 DeepMind: A New Era for AI Research
Artificial Intelligence (AI) technologies, such as Google's Gemini 2.5 Deep Think, are pushing the boundaries of complex problem-solving, leveraging advanced multi-agent parallel thinking architectures for improved reasoning depth. Deep Think's AI capabilities, notably its ability to handle multimodal inputs and long-context reasoning, are setting new standards, surpassing conventional single-agent, sequential reasoning models like those from OpenAI.
