Google has released a major update. Gemini 3 Deep Think — a platform based on reasoning artificial intelligence, designed to tackle pressing issues in science and engineering.
The development of the new version of Google Gemini 3 Deep Think was carried out in collaboration with scientists — the service is intended to address complex research tasks, which may not have clear boundaries or a single correct solution, and the accompanying data can be incomplete or unsystematic. The updated Deep Think platform is an attempt to move beyond abstract theory into the realm of practical application. The new mode is available directly in the Gemini app, but can only be accessed by Google AI Ultra subscribers; the company will also open access to the service via the Gemini API for engineers, researchers, and enterprises, which will need to submit a corresponding application.
In the Humanity’s Last Exam test, the Google Gemini 3 Deep Think platform scored 48.4% without using third-party tools; the result in the ARC-AGI-2 benchmark amounted to an unprecedented 84.6%. In the Codeforces programming test, the Elo rating was 3455; in the problems of the International Mathematical Olympiad of 2025, the system demonstrated a gold medal level, achieving the same result in solving problems in chemistry and physics competitions. In the CMT-Benchmark test, the Google Gemini 3 Deep Think mode showed a high level of understanding of theoretical physics with a result of 50.5%.
The Gemini 3 Deep Think model served as the basis for an AI agent, which in the Google DeepMind laboratory was named Aletheia. It includes a tool for verifying hypotheses expressed in natural language to identify shortcomings in proposed solutions; an iterative process for generating and refining solutions is provided. Importantly, the agent can acknowledge its inability to respond to a request. For tackling complex research, it consults Google search and web navigation tools — when preparing summaries of published literature it does not produce nonexistent references and strives to avoid inaccuracies in calculations.
Google developers have categorized the actual and potential levels of achievements of Aletheia into five grades. In a nearly fully autonomous mode, the AI agent outlined solutions for three open problems formulated by mathematician Pál Erdős—this research was rated at a zero level, corresponding to "slight novelty." In the same mode, it proposed a solution to another problem from the same set, achieving the first level or "minimal novelty." At the second level, categorized as “publishable,” Aletheia demonstrated results in autonomous mode, collaboration with humans, and as a supplementary tool. The third ("significant breakthrough") and fourth ("noteworthy breakthrough") levels have not yet been reached by the AI agent.
As per the developers' assignment, Aletheia analyzed 700 unresolved mathematical problems of Erdős and managed to solve 13 of them. However, there were already existing solutions for 9 problems, and only 4 problems appear to be genuinely solved for the first time. Moreover, out of the 212 solutions provided by the AI, only 6.5% were deemed "substantively correct." The rest either had fundamental flaws (68.5%) or solved incorrectly interpreted versions of the original problems (31.5%).
As a result, the developers agreed that their AI demonstrates "a tendency to misunderstand the question in such a way as to make it easier to answer" and remains "extremely prone to errors compared to humans." In other words, AI cannot yet replace mathematicians.
Sources:
Source: 3dnews.ru
