CSAT AI-solving Dashboard Released
Latest AI Models Top Scores, Followed by Chinese Models

Which artificial intelligence (AI) model is best at solving questions from the College Scholastic Ability Test (CSAT)? The latest models from OpenAI and Google DeepMind reportedly achieved perfect scores. Among Korean AI models, K-EXAONE 2.5, developed by LG AI Research, scored the highest.


On Nov. 20, a South Korean developer who goes by the nickname "hehee9" posted a "2027 Academic Year CSAT Large Language Model (LLM) Solving Dashboard" on GitHub, a software collaboration platform. The developer has been adding CSAT scores for AI models developed in South Korea and abroad to the dashboard since last year.


According to the dashboard, OpenAI's "GPT-6.2 Astra" and Google DeepMind's "Gemini 4 Neon" tied for first place, each scoring a perfect 450 points. Astra and Neon are the latest models from OpenAI and DeepMind, respectively.


Dashboard for solving publicly released CSAT questions. GitHub screenshot.

Dashboard for solving publicly released CSAT questions. GitHub screenshot.

View original image

They were followed by DeepMind's "Gemini 4 Argon" and OpenAI's "GPT-6.2 Sol," which scored 449.6 points each. Anthropic's "Claude Opus 5.5" and "Claude Fable 5.2" scored 449.3 points. Chinese AI companies also posted high scores. Moonshot AI's "Kimi K3.1 Pro" scored 438 points, Alibaba's "Qwen 4 Max" scored 434.5 points, and DeepSeek's "V4.1 Flash" scored 432.7 points.


The dashboard is updated by having multiple AI models immediately answer publicly released CSAT questions and then grading their responses. The AI models reportedly answer every CSAT elective subject, including all 17 inquiry subjects.


Scores for Korean AI models were also measured. K-EXAONE 2.5 ranked first among domestic models, scoring 98 points in Korean, 100 in math, and 96 in English, with little difference from leading overseas models in the core subjects. However, its converted score for inquiry subjects was only 70.7 points, widening the gap and bringing its total to 414.7 points. It was followed by Upstage's "Solar Pro 4," with 392.8 points. Krafton's "Raon-1 Mini Preview" and KT's "Mi:dm K 3.0 Pro" scored 357.2 and 352.8 points, respectively.


However, these scores are from an unofficial evaluation conducted by an individual developer. Strict comparisons may be difficult because models differ in whether they support image input and in their reasoning settings, while users' computer environments also vary. The developer said they used each model's official application programming interface (API), provided no tools such as external search or calculators when solving the questions, and entered all questions as text, attaching diagrams and graphs as images.



The developer also emphasized in an explanation of the evaluation method posted on GitHub: "The results of this test may differ from those obtained in typical usage environments on official websites or applications (apps)."


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing