GPT-5.5 vs Opus 4.7

GPT-5.5 began by analyzing the game's 30 Hz physics, coordinate systems, and player starting positions to draft a comprehensive 397-line soccer team policy. However, initial simulation testing against a "gegenpress" opponent revealed severe defensive weaknesses, resulting in three straight losses and a -2.0 average goal differential. The model quickly iterated by making a minor one-line adjustment to the policy code before successfully typechecking and submitting the final version.

Opus 4.7 began by exploring the simulation's physics, player intents, and starting positions to write an initial 11-player soccer policy. However, initial testing against a "gegenpress" opponent resulted in three straight losses and an average of seven goals conceded per match due to overly aggressive play and constant fouling. To address these vulnerabilities, the model heavily revised the policy to focus on defensive stability and reducing opponent shots before submitting the final version.