CSAT 2026

Citation frequency 1/3Korean only
What does it measure?
A fun benchmark that has AI models sit the 2026 Korean College Scholastic Ability Test (CSAT, Suneung, held November 2025). It shows what score a model would get on the same paper Korean high-school seniors take.
Who evaluates it, and how?
Subjects: Korean, math, English, Korean history and four exploration subjects (Physics I, Chemistry I, Life Science I, Society and Culture). Questions are given as OCR text (math in LaTeX) plus cropped figures, one whole section per request, through each vendor's official API with default settings and no system prompt or tools. The score is out of 450: Korean and math = common part + average of the elective parts, English (100) and Korean history (50) as is, exploration = subject average x 2. Questions and answer keys are not shown here.
Reading the score
Scores are shown as published by the source. Higher is better. Measured by an individual developer, hehee9 (CSAT-benchmark, MIT). It is a score for fun and is not used in our use-case rankings. The English listening part was replaced with its script. GPT-5.5 Pro alone was measured in the ChatGPT web app with tools disabled, not through the API.
Related benchmarks
CSAT 2027
Top score449
Models tested61
Scoring basisHigher is betterMax score 450
Last updated2026.11.17
Model release date
11 models without a release date are not plotted; they appear only in the table below.
  1. 1
    OpenAI2026.09.04 · high
    449
  2. 1
    OpenAI2026.07.09 · high
    449
  3. 3
    OpenAI
    GPT-5.5 Pro
    448
  4. 4
    Google2026.09.30 · high
    446
  5. 5
    OpenAI2026.04.23 · high
    444
  6. 5
    Anthropic2026.06.09 · high
    444
  7. 7
    Anthropic2026.09.01 · high
    444
  8. 8
    Google2026.08.13 · high
    443
  9. 9
    Anthropic2026.07.24 · high
    442
  10. 10
    Meta2026.09.02 · high
    437

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.