GPT and Gemini Achieve Perfect Scores on South Korea's College Entrance Exam

The 2027 College Scholastic Ability Test (CSAT) has concluded, revealing that major artificial intelligence (AI) models from the United States, China, and South Korea achieved perfect scores. OpenAI and Google's models both recorded a perfect score of 450 points.


The models from LG AI Research and Upstage, participating in the Ministry of Science and ICT's independent AI foundation model project, received relatively lower scores. Notably, their accuracy dropped significantly on questions that included images, indicating that their multimodal performance lags behind that of the top global models.


On November 20, Gu Yu-gyeom from Soonchunhyang University’s Department of Computer Software Engineering unveiled the '2027 CSAT Large Language Model (LLM) Solution Dashboard.'


According to the dashboard, the models that achieved perfect scores on the 2027 CSAT were OpenAI's 'GPT-6.2 Astra' and Google's 'Gemini 4 Neon.'


The Anthropic models also performed well, with 'Claude Fable 5.2' and 'Claude Opus 5.5' scoring 449.3 points each. OpenAI's 'GPT 6.1 Terra' recorded a score of 445.8 points.


Chinese models also achieved high scores. The Chinese AI company Kimi's 'K3.1 Pro' scored 438 points, receiving perfect scores in Korean, English, and Korean history but scoring 96 points in mathematics and 92 points in the inquiry section.


Alibaba Group's model 'Qwen 4 Max' also scored 100 points in Korean, English, and Korean history, while scoring 96 points in mathematics and 88.5 points in the inquiry section.


In contrast, domestic AI models showed a significant score gap compared to the top global models. LG AI Research's 'K-ExaOne 2.5' scored a total of 414.7 points, 35.3 points lower than OpenAI's Astra. It achieved perfect scores in mathematics and Korean history but scored 98 points in Korean, 96 points in English, and only 70.7 points in the inquiry section.


Upstage's 'Solar Mini 4.5 Preview' recorded 397.8 points, falling 52.2 points short of the perfect score models. It only achieved a perfect score in Korean history, with scores of 91 points in Korean, 95.7 points in mathematics, 96 points in English, and 65.2 points in the inquiry section.


Domestic models particularly struggled with image-based questions. The accuracy rate for image-included questions was 100% for GPT-6.2 Astra, Gemini 4 Neon, and GPT-6.2 Sol. Claude Fable 5.2 and Claude Opus 5.5 also reached 99.4% accuracy.


In contrast, K-ExaOne 2.5's accuracy rate for image-included questions was only 60.4%. Solar Mini 4.5 Preview recorded 52.5%, which is 39.6 percentage points and 47.5 percentage points lower than Astra, respectively.


Experts analyzed that the performance differences among these models stem from various factors, including training data, model size, and inference capabilities. Choi Byung-ho, a research professor at Korea University's AI Research Institute, noted, "There is a significant difference in the parameter size and computing infrastructure available for inference between global big tech and domestic models. Questions that include images require accurate information extraction through optical character recognition (OCR) technology, followed by inference. Since big tech has significantly improved the performance of their inference models, differences in results are inevitable."


He added, "The difference in training data is also an important factor. Global big tech companies may have had the opportunity to train on past CSAT questions and problem sets, while domestic companies may not have been able to secure sufficient related data."





* This article has been translated by AI.