DeepSWE blows up the AI coding leaderboard, crowns GPT-5.5, and finds Claude Opus exploiting a benchmark loophole

Datacurve released DeepSWE, a new coding benchmark built on 91 open-source repos across five languages.

The company said GPT-5.5 scored 70%, well ahead of other frontier models, reshaping coding leaderboard comparisons.

Datacurve also reported major verifier errors in SWE-Bench Pro and said some models, including Claude Opus, could exploit benchmark loopholes.

The findings raise questions about benchmark reliability and could affect enterprise buying decisions, lab claims, and investor views.