Google launches Gemini 3 AI for research

By Alimat Aliyeva

Google has unveiled an updated version of its Gemini 3 Deep Think, an AI platform designed to tackle complex problems in science, engineering, and advanced research, AzerNEWS reports.

Developed in collaboration with scientists, the system is optimized for research challenges where data may be incomplete, contradictory, or where no single “correct” answer exists.

The Deep Think mode is now available in the Gemini app, but full access is limited to Google AI Ultra subscribers. For the first time, the platform can also be accessed via the Gemini API, allowing engineers, researchers, and companies to integrate its capabilities into their own workflows after applying for permission.

According to Google, Gemini 3 Deep Think has achieved impressive results across multiple benchmarks. In the Humanity’s Last Exam, the system scored 48.4% without external tools. On the ARC-AGI-2 benchmark, it achieved 84.6%. In programming tests on Codeforces, the model earned an Elo rating of 3455. In the 2025 International Mathematical Olympiad tasks, it performed at a gold medal level, with similar high marks in chemistry and physics Olympiads. On the CMT-Benchmark for theoretical physics, the system scored 50.5%.

Based on Gemini 3 Deep Think, Google DeepMind developed an AI agent called Aletheia. Aletheia is capable of analyzing scientific hypotheses, identifying flaws in reasoning, and suggesting refined solutions. Notably, the agent can recognize when a problem is beyond its capabilities and indicate that no answer can be provided.

To tackle complex research, Aletheia leverages Google search tools and web navigation. When reviewing scientific literature, it minimizes errors and avoids citing nonexistent sources — a frequent problem for AI in academic contexts.

Developers evaluated the agent’s performance across five levels of autonomy. In fully autonomous mode, Aletheia proposed potential solutions to three open problems posed by renowned mathematician Paul Erd?s, though these were classified as “minor novelties.” Among 700 unresolved Erd?s problems analyzed, it suggested solutions for 13; nine of these had already been solved, leaving only four genuinely new contributions. Of 212 proposed solutions overall, only 6.5% were considered “meaningfully correct,” with the remainder containing errors or misinterpreted problem statements.

Developers acknowledge that the system still tends to simplify problem formulations to increase the likelihood of finding an answer, and it remains more error-prone than human experts.

Despite these limitations, Gemini 3 Deep Think and Aletheia are being hailed as promising tools for accelerating scientific research. Experts believe that while AI is not yet capable of replacing professional scientists or mathematicians, it can serve as a powerful assistant — helping researchers explore hypotheses, test ideas, and sift through massive amounts of data far more efficiently than before.

Some futurists even speculate that, over time, agents like Aletheia could become collaborators in frontier research, helping humans tackle problems once considered too complex or computationally intensive to approach.