Those who pay $250 per month for the Gemini Ultra subscription can now use Gemini 2.5 Deep Think. The so-called multi-agent AI model is designed to solve complex problems by using multiple thought processes simultaneously.
Starting this Friday, Google DeepMind is rolling out Gemini 2.5 Deep Think to high-paying users. The model costs $250 per month via the Ultra subscription. The system works differently from traditional AI models by deploying multiple AI agents in parallel on a single question.
This approach requires much more computing power, but according to Google, it delivers significantly better results. The company demonstrated this by winning a gold medal at the International Math Olympiad with a variant of the model. That version takes hours to perform complex calculations, rather than the seconds or minutes required by conventional AI models.
Google claims that Gemini 2.5 Deep Think leaves the competition behind. On Humanity’s Last Exam, the model scores 34.8 percent without tools. xAI’s Grok 4 scores 25.4 percent and OpenAI’s o3 remains stuck at 20.3 percent. The difference is even greater in programming challenges. Gemini 2.5 Deep Think achieves 87.6 percent on LiveCodeBench6. Grok 4 achieves 79 percent and o3 72 percent.
Multi-agents
Several AI players are experimenting with multi-agent systems. Elon Musk’s xAI recently launched Grok 4 Heavy, also based on this approach. OpenAI is also said to be using a multi-agent model with its suspected GPT-5. Anthropics Research agent also works with multiple AI components for thorough research reports. The higher costs are currently limiting these systems to expensive subscriptions. This is not surprising: a single multi-agent task is essentially the sum of tasks performed in parallel.
Google plans to conduct further tests via the Gemini API in the coming weeks. The company wants to discover how developers and businesses can use the multi-agent system for practical applications.