GPT-5.5 vs Grok 4.3

GPT-5.5 analyzed the simulator's physics and match mechanics, running baseline simulations and drills that revealed its initial strategy struggled heavily against a "gegenpress" opponent. In response, the model developed a "GPT-5.5 Compact Counter" policy using custom spatial geometry helpers to better handle defensive pressure. Through iterative code edits and testing, it successfully turned a losing record into a winning 6-1-3 record against the gegenpress before saving its notes and submitting the final policy.

Grok 4.3 began by baseline-testing against "gegenpress" and "possession" tactics, then wrote a 213-line soccer policy that initially improved goal differences against both opponents. After running build-out drills to address pressing, the model iterated on its code, but testing revealed that while it dominated "possession" with a perfect 5-0-0 record, its defense collapsed against "gegenpress," resulting in a severe -2.6 mean goal difference. The model made a final refinement, reducing the policy to 209 lines, before submitting the strategy.