Gemini 3.5 Flash vs Opus 4.8

The model began by analyzing the game's physics and player action space to write an initial 283-line team policy, which suffered five consecutive losses against a "gegenpress" opponent. To improve, it ran detailed analyses on the engine's set-piece rules, boundary conditions, and match statistics to refine its approach. However, its updated 381-line policy performed worse, conceding an average of 9.2 goals per match, leading the model to investigate shot speed thresholds before the session timed out.

Opus 4.8 began by analyzing the 30 Hz arcade physics engine and player coordinates to write an initial positional soccer policy. When simulations against a "gegenpress" opponent resulted in a losing record and three conceded goals per game, the model implemented a dual-presser marking system, which initially backfired by causing excessive fouls and a worse goal differential. To fix this, the model rewrote the policy to reduce tackle commitment and restrict the second chaser's tackling, ultimately securing a much-improved neutral goal difference and reducing conceded goals to 1.2 before final submission.