Google DeepMind has officially unveiled its new artificial intelligence model, Gemini 2.5 Deep Think, which the company describes as its most advanced system for complex reasoning. Developers emphasize that this model employs a multi-agent approach: simultaneously testing multiple hypotheses to find optimal solutions for the most challenging tasks.
This is reported by Finway
Premium Access and Unique Features of Gemini 2.5 Deep Think
The first to take advantage of Gemini 2.5 Deep Think will be subscribers to the Google Ultra plan, which costs $250 per month. Starting this Friday, users of this plan will gain access to the new model directly within the application.
Advantages of the Model Over Competitors and Practical Applications
The presentation of Gemini 2.5 Deep Think first took place at the Google I/O 2025 conference. According to company representatives, this innovative system helped Google win a gold medal at the International Mathematical Olympiad. At the same time, a limited version of the model, adapted for tasks requiring prolonged reasoning and aimed at scientific research, is available to researchers.
“The company claims that Gemini 2.5 Deep Think has been significantly improved compared to the spring version. Google developers have implemented new reinforcement learning methods, which have enhanced the effectiveness of reasoning algorithms.”
In the HLE test, which assesses the AI’s ability to answer questions across various fields of knowledge, Gemini 2.5 Deep Think scored 34.8% without additional tools. In comparison, Grok 4 from xAI received 25.4%, while the o3 model from OpenAI scored 20.3%. In LiveCodeBench6 — a programming test — Gemini 2.5 Deep Think also came in first place with a score of 87.6%, surpassing Grok 4 (79%) and o3 from OpenAI (72%).
Google notes that the new model is ideally suited for solving tasks related to creativity, strategic thinking, and implementing complex solutions. The company hopes that feedback from researchers will help refine the multi-agent architecture for further applications in science and other fields.