There appears to be little doubt now that AI is already better than most of the world’s best programmers.
The AtCoder World Tour Finals Heuristic 2026 wrapped up in Tokyo with OpenAI’s model finishing so far ahead of the 12 human finalists that even the contest’s own organizer described himself as “completely defeated.”
The result closes the loop on a rivalry that started last July, when OpenAI’s model lost the same event by the narrowest of margins to a human competitor named Psyho, who later admitted he’d survived on barely any sleep during the contest week. This year there was no last-minute drama. Psyho, who wasn’t among the 2026 finalists himself but followed the contest closely and posted commentary throughout, summed it up bluntly on X once the standings were finalized: “Humanity has not prevailed.” He added that humans had “performed quite well” overall, but that OpenAI’s performance was “outstanding,” and that he’d publish a longer writeup on the contest in the coming days since it deserved more than a quick post.
Chokudai, AtCoder’s founder, competed in the exhibition himself and posted a reaction after the results came in. He called the problem “a really great” one that showcased human creativity, and said that despite having plenty of contest time available, he still ended up “utterly overwhelmed by AI.” He framed the loss as a milestone rather than a defeat for the field as a whole, arguing that a machine capable of overwhelming the strongest heuristic competitors on the planet points toward AI eventually solving a volume of problems “beyond imagination.”
The format itself was built for exactly this kind of showdown. The AWTF Heuristic finals ran across two days in Tokyo, with 12 of the world’s best heuristic programmers qualifying based on their results across AtCoder’s contests throughout 2025. Competitors were given a single, deliberately unique optimization problem, the kind where there’s no known best answer and the goal is to get as close to optimal as possible within the time limit. Day one results were kept hidden from the public to avoid spoiling the contest, with the full leaderboard and a live ten-hour broadcast reserved for day two. AtCoder had also set up a special “Humanity Prevails Award” for whichever human competitor managed to beat OpenAI’s model outright, a prize that went unclaimed this year.
The final scores made the gap explicit. OpenAI’s submission finished with a score in the tens of billions, more than seven times higher than the best human result. Based on Psyho’s list of congratulations, Shun_PI appears to have led the human field, with ALGO ARTIS’s terry_u16, Preferred Networks’ yokozuna57 and asi1024, and researcher yosupo among the other finalists rounding out the standings.
This isn’t the first time an AI system has beaten strong human competition on an AtCoder heuristic problem. Sakana AI’s ALE-Agent placed first outright in an AtCoder Heuristic Contest against more than 800 human entrants earlier this year, though that was a shorter, open-format event rather than an invite-only finals with the world’s top specialists. What makes the 2026 AWTF result different is the caliber of the opposition. These are competitors who spent all of 2025 grinding through qualifying rounds to earn one of just 12 seats at the table, and OpenAI’s model still put daylight between itself and every one of them.
The win also lands at a point where OpenAI has been racing to keep its coding models ahead of rivals on other fronts. The company shipped GPT-5.5 in April with claims of state-of-the-art coding performance at a lower cost than competing frontier models, and followed it with the GPT-5.6 series in June, explicitly positioned against Anthropic’s newer Mythos-tier models on agentic coding benchmarks. AtCoder gives the company a much more public proving ground than a benchmark chart, one where a model has to hold up against elite human problem-solvers live, on camera, with no do-overs.
For competitive programmers, the takeaway from Psyho’s running commentary during the contest is worth sitting with. He noted that watching an AI’s superior score sit at the top of a live leaderboard tempts human competitors to abandon reasonably good solutions in search of some perfect approach that may not exist, often leaving them with too little time to implement anything at all. Whether that’s a fair account of what happened in Tokyo or not, the scoreboard suggests the gap between the best human heuristic programmers and OpenAI’s model has widened rather than narrowed since last summer, and by a considerable amount.