REASONING

OpenAI's System Swept Both Divisions at AtCoder's Finals

At the AtCoder World Tour Finals in Tokyo, an OpenAI system beat all 12 human finalists in both the Algorithm and Heuristic divisions — a year after losing the same event by the narrowest margin.

The AtCoder World Tour Finals is an annual invitational that gathers the twelve competitive programmers who qualified through the year's official AtCoder contests, flies them to Tokyo, and runs them through two separate competitions: an Algorithm division of discrete problems with exact answers, and a Heuristic division built around a single open-ended optimization problem with no known best solution. Since 2025, AtCoder has run an unofficial 'Human vs. AI' exhibition alongside the human standings, entering a system from OpenAI in both divisions without counting its score toward the official ranking.

This year, the AI won both. In the Algorithm division, held July 9, 2026, over a seven-hour window, OpenAI's system solved all five problems for 8300 points. The best of the twelve human finalists, a competitor racing under the handle tour1st, solved three for 4300. Two of the five problems, labeled C and E, were solved by none of the twelve humans. In the Heuristic division, held July 7-8, the system's score came in at the tens of billions on the contest's scale, more than seven times the best human result, posted by the human leader, Shun_PI.

The result reverses what happened at the same event a year earlier. In the 2025 Heuristic division, OpenAI's system led for most of the ten-hour contest before Przemysław Dębiak, competing under the handle Psyho, overtook it in the closing stretch to win by the narrowest of margins; Dębiak later said he had run on almost no sleep to hold the lead. In the 2025 Algorithm division, OpenAI's models were not competitive at all. Twelve months on, both gaps closed and reversed.

AtCoder had set aside a 'Humanity Prevails Award' of 600000 yen, on top of the standard 1000000 yen first prize, for any human finalist who both beat the AI system and finished first overall. It went unclaimed in both divisions this year, as it had in the Algorithm division the year before. Dębiak, competing this year as a finalist rather than as the reigning champion defending a win, wrote afterward that 'humanity has not prevailed' and called the system's performance outstanding.

A member of OpenAI's team, Borys Minaiev, described the system to reporters as close to GPT-5.6, run with an extended test-time compute budget rather than any specialized training for competitive programming. He said the team had tested it against the previous years' World Tour Finals problems beforehand, and it had solved all of them, usually in under an hour — a sharp contrast with problems D and E in this year's live Heuristic contest, which he said were harder than anything the team had seen the system face before, with problem D alone taking roughly three hours.

The World Tour Finals measures a narrower skill than most of what this beat covers: solving discrete, well-specified problems and open-ended optimization problems under a strict clock, not playing another agent directly. But the shape of the result is familiar from elsewhere in machine competition this year — a benchmark that looked like a genuine human stronghold in one cycle gets cleared in the next, sometimes by a margin wide enough that the contest's own escape clause goes unused.