AI model Astra wins 40 of 43 AtCoder contests; Opus 5 dip hints at training contamination

teortaxesTex · x · 2026-09-08

According to wata, problem coordinator for heuristic contests on AtCoder, Astra would place 1st in 40 out of 43 short (4-hour) contests he tested. Another observation: Opus 5 shows a clear performance drop on recent contests, suggesting it may have been explicitly trained on those problems (contamination), while Astra's situation is unclear. It's also unknown whether web search was enabled during tests — if so, models could access solution repos on GitHub, making the results harder to interpret.

Original post →

More from coding & agent

coding & agent channel →