OpenAI 模型在 AtCoder 世界巡回赛决赛击败顶尖人类程序员

But human programmers aren’t obsolete — at least not yet.

The AtCoder World Tour Finals, held in Tokyo every year, is one of the most prestigious programming competitions in the world. It has two divisions. There’s a heuristic division where programmers compete to maximize performance on an open-ended task. And there’s an algorithmic division where contestants must find a way to efficiently compute exact solutions to mathematically challenging problems.

During last year’s competition, Polish programmer Przemysław Dębiak (known as “Psyho”) narrowly claimed first place in the heuristic division. He beat 11 human competitors — and an internal OpenAI model trained to be especially strong at reasoning tasks.

“Humanity has prevailed (for now!)” he wrote in a tweet right after the competition. OpenAI’s model came in second after leading for most of the 10-hour competition, a surprisingly strong result for AI models at the time.

OpenAI’s models last year weren’t good enough to compete in the algorithmic division.

The 2026 competition, held this week, turned out very differently. Organizers chose a heuristic problem designed to help humans succeed. Despite that, OpenAI “completely demolished human competitors,” Psyho noted after the two-day competition finished Wednesday night. It’s hard to quantify exactly how big the AI’s margin of victory was, but Psyho told me that he would guess that humans would need to work at least a few more days to match the AI’s score — though he stressed that this is a hard number to predict exactly.

The next day, OpenAI’s system crushed humans on the algorithmic problems as well. Over the course of the seven-hour competition, it solved all five problems, including two that none of the 12 human competitors — all among the best in the world — were able to solve.

So at the award ceremony for the 2026 AtCoder competition, the organizers presented two “humanity surrenders” awards to OpenAI for its models’ performances in the two competitions.

This was probably the last time humans had a realistic shot at winning a programming competition against top AI models. Today’s AI models can find impressive, elegant solutions much more quickly than humans. And future models will only get better.

In some ways, OpenAI’s performance was even more impressive than the raw score suggests.

Keep reading with a 7-day free trial