AI Models Excel in 2027 College Entrance Exam Simulation


According to results from the 2027 College Scholastic Ability Test (CSAT) simulation conducted on major domestic and international artificial intelligence (AI) models, nine out of 28 evaluated models achieved perfect scores across all four areas: Korean, mathematics, English, and Korean history. Models from American companies OpenAI, Google, and Anthropic dominated the top ranks, with Chinese AI closely following. Among domestic models, LG AI Research's 'K-EXAONE 2.5' received the highest score.

On November 19, the day after the CSAT, the AI benchmark platform 'Next CSAT' input the exam questions for evaluation and scoring. The scoring system was based on a total of 426 points, with 124 points for Korean, 152 points for mathematics, 100 points for English, and 50 points for Korean history. The scores for Korean (speech and writing, language and media) and mathematics (probability and statistics, calculus, geometry) included all elective subject scores.

A total of nine models achieved a perfect score of 426 points. By developer, these included: ▲ Anthropic's 'Claude Fable 5.2' and 'Claude Opus 5.5' ▲ Google's 'Gemini 4 Argon, Carbon, Hydrogen, Neon' ▲ OpenAI's 'GPT-6.1 Terra' and 'GPT-6.2 Astra, Sol'.

Several top models narrowly missed perfect scores. Anthropic's 'Claude Sonnet 5.5' scored 422 points (Korean 122, mathematics 152, English 98, Korean history 50), while 'Claude Haiku 5.5' scored 420 points (Korean 120, mathematics 152, English 98, Korean history 50).

The rise of Chinese AI models was also notable. Moonshot AI's 'Kimi K3.1 Pro' lost only 4 points in mathematics, achieving a total of 422 points (Korean 124, mathematics 148, English 100, Korean history 50), while Alibaba's 'Qwen4 Max' also scored 420 points, including perfect scores in Korean and English (Korean 124, mathematics 148, English 100, Korean history 48).

Among lightweight models, OpenAI's 'GPT-6 Luna' scored 416 points (Korean 122, mathematics 149, English 100, Korean history 45), and Google's 'Gemini 3.5 Flashlight' scored 384 points (Korean 116, mathematics 128, English 95, Korean history 45).

Among domestically developed AI models, LG AI Research's 'K-EXAONE 2.5' topped the rankings with a total of 368 points across three areas (Korean, mathematics, English) (Korean 120, mathematics 152, English 96). Notably, it achieved a perfect score in the mathematics area, including all elective subjects.

Following were ▲ Upstage's 'Solar Mini 4.5 Preview' with 352 points (Korean 111, mathematics 145, English 96) ▲ Upstage's 'Solar Pro 4' with 346 points (Korean 109, mathematics 149, English 88) ▲ Krafton's 'Raon-1 Mini Preview' with 338 points (Korean 101, mathematics 146, English 91) ▲ KT's 'Mideum K 3.0 Pro' with 317 points (Korean 74, mathematics 149, English 94). Domestic models generally showed strength in mathematics, while there was significant variation in the Korean area among models.

However, some domestic models did not have scores for Korean history, so only the total scores for three areas (with a maximum of 376 points) were disclosed. Therefore, there are limitations in comparing total scores with the 426-point models that include Korean history.

These scores are based on a self-assessed raw score system that aggregates all elective subjects, which differs from the standardized scores displayed on actual test-taker score reports. Additionally, differences in execution settings and input methods among models make it difficult to definitively assess AI performance based solely on these figures.




* This article has been translated by AI.