AI is getting better at breakneck pace, and it’s now managed to do better than the best humans writing the prestigious IIT JEE exam.
Google’s latest Gemini 2.5 Pro model has scored better than the candidate getting All India Rank 1 in the IIT Jee Advanced Exam for 2025. The top scorer in the IIT JEE exam was Rajit Gupta, who’d scored 332/360 and bagged the top rank. Google’s Gemini 2.5 Pro model did slightly better than him, managing 336/360, which meant that the AI model did better than every single human being that wrote the exam.
The IIT JEE exam is one of the hardest college exams in the world. It has two stages — the first with nearly 10 lakh candidates, and another stage with the top 2.5 lakh candidates from the first exam. The exam has 3 parts of 2 hours each, with 3 subjects — Physics, Mathematics and Chemistry. Each correct answer is worth 3 marks, while 1 mark is deducted for every wrong answer.
On the test, administered by Bytedance, Gemini 2.5 Pro topped the AI models that were tested. Gemini scored the highest in mathematics with a score of 119.6. It scored 108.6 in Physics and 108 in Chemistry. Bytedance’s own model, the Seed 1.6, scored 329.6 overall, and would’ve ranked 4th among all candidates that wrote the exam. Claude Opus 4 scored 314.4, and would’ve secured an AIR of 13. OpenAI’s o4 mini high scored 308.4, and would’ve ranked 18th in India.
These are pretty incredible results. The IIT exam is notoriously difficult, and students prepare for years to be able to get a good rank. The exam doesn’t simply rely on rote memorization, but tests concepts and critical thinking skills. AI systems, though, are now performing better than the best human candidates in the exam. And with AI systems already better than the best humans that are aged 18 and about to enter college, it might not be long before they’re better than the best humans at the post graduate and PhD level. And this intelligence — which is near-instant and available for nearly free — could change our society in ways that are hard to conceive at the moment.