South Korean AI Startup Posts Near-Perfect Score on JEE Advanced 2025 Exam Paper

Artificial Intelligence Startup Claims Breakthrough in Academic Contexts

A South Korean artificial intelligence startup, backed by OpenAI, has announced that its system has achieved a score equivalent to 351 out of 360 on problems drawn from JEE Advanced 2025 — one of the most demanding engineering entrance examinations in India. This development has reignited debate about the reliability and transparency of AI reasoning in academic contexts.

The startup, GPAI, attributes the result to an architecture that pairs large language model-based reasoning with a deterministic computational engine. According to GPAI, this combination produces solutions that are more structured and easier to verify than those generated by conventional AI systems. Additionally, the model incorporates visual interpretation, allowing it to work with diagrams, a capability potentially relevant to subjects such as mechanics, optics, and calculus, where figures are often integral to problem-solving.

This achievement comes after a widely discussed experiment by Anushka Aashvi, a student at IIT Kharagpur, who tested OpenAI's ChatGPT o3 on a JEE Advanced 2025 mock paper. To simulate actual exam conditions, she prompted the model to approach the questions as a JEE aspirant would — without web access, coding tools, or any hints. Each question was presented in a separate chat session to eliminate memory carry-over, and no corrections were offered during the process. ChatGPT o3 scored 327 out of 360 under those conditions, a tally that would have placed it at All India Rank 4 in the actual examination.

Read also: AI Visionaries Support Startup Developing Event Prediction Models

GPAI's system approaches problems by decomposing them into smaller logical steps before arriving at a final answer — a method the company argues makes it more suitable for academic environments where the reasoning process carries as much weight as the outcome. The startup claims its platform has attracted users across the United States, Europe, and Asia since its launch, with notable uptake in India. Specifically, it has accumulated over 1,000 users from IIT Delhi alone within three months of going live, and reports broader interest from students and researchers at various IIT campuses.

The growing performance of AI systems on high-stakes academic benchmarks has shifted the conversation among educators from whether such tools can produce correct answers to whether they can do so in ways that are transparent and dependable enough for serious scholastic use. GPAI has published demonstrations and sample problem-solving workflows on its platform for those who wish to examine its methodology more closely.