Paper Says AI 'GrandCode' Took First in Three March Codeforces Rounds, but Official Standings List Humans

A revised paper posted to arXiv claims an artificial intelligence system called GrandCode won three live Codeforces programming contests in March. But Codeforces’ own official standings for those same rounds, in the platform’s default view of “trusted participants only,” list human winners instead.

The paper, “GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning,” was updated July 13 by the DeepReinforce Team. It says GrandCode “placed first” in Codeforces Round 1087 on March 21, 2026, Round 1088 on March 28 and Round 1089 on March 29. That makes the result notable, but also unresolved: the claim is not independently confirmed by Codeforces’ official trusted standings.

According to the author-hosted PDF, GrandCode competed under three different handles: averyjones1 in Round 1087, yokeko in Round 1088 and Vortex1 in Round 1089. The paper also reports two scoring methods. One, called S(separate), adds up scores at the time each solution was completed. The other, S(joint), gives a score for the full set of submissions in a single account. That distinction matters. In Round 1088, for example, the paper reports an S(separate) score of 16,511 and an S(joint) score of 15,008. Codeforces’ trusted-participant winner for that round, turmax, is listed at 15,677. In other words, GrandCode’s “first place” claim in that round appears to depend on the paper’s custom S(separate) method, because its own S(joint) score falls below the official trusted winner’s score.

The same gap appears when comparing the paper’s claims with Codeforces’ default standings pages. For Round 1087, the paper reports 9,269 under S(separate) and 8,334 under S(joint), while the trusted standings show bismispis in first place with 8,057. For Round 1088, the trusted winner is turmax with 15,677. For Round 1089, the paper reports 11,596 under S(separate) and 9,506 under S(joint), while the trusted standings show NotNanYan on top with 8,950. Codeforces is a major live competitive-programming platform, and its public standings distinguish between trusted participants and other accounts that can appear if users choose the “show unofficial” option instead of the default trusted-only view.

That split helps explain why the paper’s screenshots may not match the official standings page most readers first see. The research found that third-party contest archives and aggregators such as CLIST appear to reflect snapshots that include unofficial participants and show the handles cited in the paper. But the research did not identify a public statement from Codeforces saying GrandCode’s results were recognized as official contest victories.

The paper itself acknowledges a key complication. “It is worth noting that Codeforces has policies against AI-generated content, and accounts suspected of using AI face removal. High-ranking accounts in the contest are under even tighter scrutiny,” it says. The abstract makes the broader claim in stronger terms, stating: “GrandCode is the first AI system that consistently beats all human participants in live contests of competitive programming: in the most recent three Codeforces live competitions, i.e., Round~1087 (Mar 21, 2026), Round~1088 (Mar 28, 2026), and Round~1089 (Mar 29, 2026), GrandCode placed first in all of them, beating all human participants, including legendary grandmasters.”

If verified as official wins, that would mark a significant step beyond earlier AI coding benchmarks. The paper says the strongest comparison point it cites, Google’s Gemini 3 Deep Think, reached eighth place and was not tested under live competition conditions. But for now, the central point is narrower: a major new AI paper says its system topped three live Codeforces rounds, while Codeforces’ own default official standings for those contests still show human winners.