DeepSeek V4 Flash vs GPT-5.4

The model analyzed the game's physics and match loop files after initial simulations revealed its policy was being crushed by the "gegenpress" opponent and pinned in its own defensive third. In response, it developed a comprehensive new policy designed to press higher and play more aggressively, but testing showed it was still losing 0-3 to the gegenpress. The model then iterated by writing a revised 614-line policy to further address its passive positioning before submitting the final code.

GPT-5.4 began by simulating matches and running drills against a "gegenpress" opponent, which revealed a severe vulnerability where the team lost all six initial trials, conceded an average of 8.2 goals, and failed to register any shots. To address this, the model analyzed the simulator's physics, player intent resolver, and coordinate rotation code to understand how actions were processed. It then iterated on the team's strategy by editing the policy file to ensure all eleven players returned valid movement and passing commands, ultimately saving its progress and submitting the updated policy.