GPT-6 Astra achieves a perfect score on South Korea's college entrance exam

OpenAI's latest model, GPT-6 Astra, achieved a flawless score on an evaluation based on South Korea's college entrance examination, according to benchmark data posted on GitHub. The test covered subjects ranging from advanced mathematics and sciences to history and literature without connecting to the internet. As reported by The Korea Times, Astra is the first system to get every question right across the full suite of subjects.

What sets this performance apart is how efficiently the software worked. Log files show that Astra used far fewer reasoning tokens than competing systems built by Anthropic or Google, as well as earlier OpenAI models. An academic expert quoted by The Korea Times noted that the software saves computational resources by retaining and compressing its previous reasoning steps instead of starting over for every task.

While impressive, industry experts warned that performing well on standard tests does not mean an AI can handle complex workplace assignments. Additionally, because test materials were already public, it is unclear if prior exposure in training data played a role. The ultimate test will be whether this context-saving design helps models solve genuinely new problems that lack standardized answer keys.