In the recently finished Round 1112, I somehow got retked by a Div. 1B for two full hours for no apparent reason. I knew my approach was right, but there was some very very tiny bug somewhere in my code.

Wait, why not ask GPT to debug the code, quickly move on to the next problem, and avoid having my entire evening fxxked? The thought flashed through my mind, but instead, it inspired a more interesting question: when faced with a long and painful debugging session, and the option of using “a little magic,” how many people would choose the latter?

Since GPT can debug almost any problem, if many people choose the latter, then we should be able to observe that once something goes wrong with a problem during a contest, the probability of eventually solving it becomes much higher than it was before GPT.

So I asked my Codex to collect submissions from recent years and calculate how the debugging success rate has changed each year. Here, the success rate is defined as follows: after a submission receives a WA, RE, or another failed verdict, whether the problem is eventually solved before the contest ends.

The result was shocking. In the pre-GPT era, this value remained stable at around 53%. It began to rise in 2025—GPT-5 was out—and reached an astonishing 69.0% in 2026.

This means that if we assume the success rate of purely human debugging is 50%, while GPT’s success rate is 100%, then approximately 38% of contestants who face with difficulties ask GPT to “help a little.”

Of course, this statistical result may not be particularly professional. I hope someone else can reproduce it and produce a more rigorous report. But for now, it seems that the entire competitive programming community has gradually learned to “coexist” with GPT.