Opus 4.8 vs MiMo v2.5 Pro

Opus 4.8 began by simulating matches and running defensive drills against a "gegenpress" opponent, which revealed that its team was getting pinned in its own third due to passive ball-carrying. In response, the model edited the policy to make ball-carrying more progressive and improve support spacing. This adjustment successfully turned a losing record into a 3W-1D-1L victory against the press while maintaining a dominant 4W-1D-0L record against possession-based opponents before final submission.

MiMo v2.5 Pro began by analyzing the soccer engine's physics and command systems, running baseline simulations that revealed a severe vulnerability to high-pressing opponents. In response, the model wrote a 446-line policy designed to improve chance creation and resist defensive pressure. While subsequent testing showed the new strategy successfully dominated possession-based opponents with a 4W-1D-0L record, it still struggled to overcome the gegenpress, leading the model to submit the partially improved policy.