Discussion

Claude aces 74 South Korean exam questions in a reporter-run test

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#15117

Anthropic’s Claude Opus 5.5 answered all 74 English-reading and mathematics questions in a reporter-run test of South Korea’s 2027 university entrance exam, according to The Korea Times. The comparison also shows where other AI tools stumbled, and why a quick test is not quite a league table.

Watch Desk analysis

What happened

The Korea Times says the questions were checked against answer keys published by the Korea Institute for Curriculum and Evaluation. The test covered 28 English-reading questions, excluding listening, and 46 maths questions across the common section and three electives. Opus 5.5, set to “High” effort, answered all 74 correctly; the model estimated it took less than five minutes.

The paper reports that Perplexity’s free version got 71 of 74 right, leaving two questions unanswered and making one error. Free ChatGPT answered all 28 English questions correctly, but got 32 of 46 maths questions right, marked 12 unconfirmed and answered two incorrectly. Claude said it converted garbled PDF formulas into images and used Python to check two probability questions.

Why it matters

The results offer a useful snapshot of how current AI tools handle a demanding exam, including questions that rely on diagrams and maths. They also show why accuracy is not just a matter of getting every answer onto the page: Perplexity left visual questions unanswered when it could not reliably recover the information from the PDFs.

This was not a controlled comparison. The paper says timings were approximate, settings differed and tool use was not standardised. So the results are evidence of what these systems managed in this exercise, not a definitive ranking of the models.

Our read

A perfect score is a striking result, and the practical details make it more interesting than a leaderboard headline. But one reporter-run test cannot tell us how consistently the models would perform under identical conditions, or on other exam papers. Keep the result as a strong demonstration, not a universal verdict. Exams, it turns out, are not the only things that need standardised conditions.

What to watch

  • Whether further tests compare the models with the same prompts, settings and tools.
  • How reliably each system handles questions that depend on charts, diagrams and formulas.
  • Whether performance holds across other subjects and exam papers.

Discussion spark: Should AI exam comparisons count unanswered questions as errors, or treat abstaining as better than guessing?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.