An AI Agent Won a Rated Coding Contest Outright, for the First Time

On December 14, 2025, AtCoder ran Heuristic Contest 058, a four-hour rated event open to anyone with an account: solve one optimization problem, submit as many times as the clock allows, get ranked by score against everyone else who entered. Sakana AI, a Tokyo-based lab, entered its own system under the handle fishylene. It took the lead about two hours in and held it to the end, finishing first among 804 participants. Second place went to a human competitor racing under the name yosupo.

The system was ALE-Agent, built by running several frontier language models in parallel to generate candidate solutions, testing them, and feeding the results back in for another round, repeated for the full four hours. Sakana has said the run consumed roughly $1,300 in compute across more than 4,000 model calls. That is a meaningful sum for one contest entry, and also the kind of number that only makes sense to spend if the alternative is a team of engineers doing the same search by hand.

The problem itself concerned hierarchical production planning — arranging a set of interdependent manufacturing steps, each with its own constraints, into an efficient schedule, a shape of problem that shows up in real supply chains as well as contest sets. According to Sakana, the intended approach was a greedy construction followed by standard simulated annealing. ALE-Agent instead converged on a different heuristic, layering a custom scoring function it derived during the contest on top of a simulated-annealing search with more varied move types than the reference solution used. It reached a better answer by searching a differently shaped space, not by running the expected method faster.

That distinction is the more useful part of the result. A system that wins by brute-force search over a known-good method is a statement about compute. A system that wins by finding a different method than the one the problem setter had in mind is a statement about the search itself — the four hours were spent partly on developing an approach, not only on tuning one.

This is a narrower claim than it might sound. AHC058 is one rated contest among the roughly monthly heuristic events AtCoder runs, not the annual World Tour Finals invitational that draws the year's top twelve qualifiers to Tokyo — the venue where, in July 2025, an OpenAI system led an equivalent optimization contest for most of ten hours before Przemysław Dębiak overtook it in the closing stretch to win by a narrow margin, and where, a year later, an OpenAI system swept both of that event's divisions outright. ALE-Agent's win at AHC058 sits earlier and lower-stakes than either of those: a regular season contest, not a finals, decided by margin over 804 entrants rather than twelve.

Placed against that timeline, the shape is consistent rather than new: AI systems have gone from finishing off the podium, to finishing second against a small elite field, to winning a large-field rated event outright, over roughly a year and a half of AtCoder's own contest history. What AHC058 adds is a case where the winning agent did not just execute known technique under time pressure, but arrived at a heuristic the contest's own problem setter had not anticipated.