Google’s Gemini 2.5 pro outperforms humans and other AI models in JEE Advanced!
Google's advanced Artificial Intelligence (AI) model, Gemini 2.5 Pro, has achieved the highest score in India’s Joint Entrance Examination (JEE) Advanced 2025, one of the toughest college entrance tests globally.
It has outshined over a million human aspirants and surpassed several leading AI competitors. This was revealed in a study published by Chinese tech giant ByteDance last week, reported Outlook Business.
According to the technical report, Gemini 2.5 Pro secured 336.2 out of 360, placing it at the top of the leaderboard. In comparison, the top human scorer achieved 332, while the 10th-ranked human candidate scored 317.
"Compared with the all-India human test takers, the first place scored 332 points and the 10th place scored 317 points. Gemini-2.5-Pro and Seed 1.6-Thinking were able to achieve top 10 scores in India. Gemini-2.5-Pro performed well in Physics/Chemistry, and Seed 1.6-Thinking answered all the 5 sampled math tests correctly," ByteDance stated in its report.
The study highlights the rapidly advancing capabilities of large language models (LLMs) in solving complex academic assessments with increasing accuracy.
The JEE Advanced, considered one of the world’s toughest academic entrance exams, is taken only by the top 2,50,000 students who clear the JEE Main. It comprises two rigorous three-hour papers in Mathematics, Physics, and Chemistry, with negative marking that demands high levels of accuracy and precision.
Among other high-performing AI models, Anthropic’s Claude Opus 4 scored 314.4, earning the 13th spot, while OpenAI’s o4-mini-high followed close behind in 18th place with a score of 308.4.
Seed 1.6 - Thinking, an enhanced version of ByteDance’s earlier model —showed improved capabilities through upgraded computational power and a larger training dataset that includes mathematics, programming, puzzles, and visual content. It employs long-chain-of-thought reasoning to tackle complex problems, although this can occasionally lead to overthinking.
To further sharpen its reasoning, Seed 1.6 - Thinking employs parallel decoding, enabling it to handle more tokens before generating an answer. This strategy enhances performance on challenging problems without requiring retraining. Notably, the model improved by 8 points on the challenging Beyond AIME math test and showed significant strides in coding tasks as well.
The success of such models is fueling the AI industry’s push toward advanced reasoning systems, which are often described as “superintelligence.”
Reflecting this momentum, Meta CEO Mark Zuckerberg recently announced the formation of Meta Superintelligence Labs, bringing onboard high-profile AI leaders such as Alexandr Wang (former Scale AI CEO) and Nat Friedman (former GitHub CEO).
The move has intensified the ongoing war in Silicon Valley. Reports suggest that Meta is actively recruiting AI researchers, including Shengjia Zhao, Jiahui Yu, Hongyu Ren, and Shuchao Bi from OpenAI.