LLM Debate Benchmark adds GPT-5.5, Grok 4.3

This is a useful reminder that adversarial debate is a different skill from raw chat quality: the leaderboard is compressing in the middle, but the newest names are not automatically the strongest. The side-swapped setup and three-model judging make the result more credible than a single-pass preference test, even if judge agreement is only moderate.

• –GPT-5.5 entering below GPT-5.4 is the sharpest signal here: newer model, weaker debate showing
• –Grok 4.3 underperforming Grok 4.20 Beta 0309 suggests a real regression, not just leaderboard noise
• –GLM-5.1, Kimi K2.6, DeepSeek V4 Pro, and Xiaomi MiMo V2.5 Pro all look like solid incremental gains, but not a frontier break
• –The benchmark’s side-swapped format matters because motion asymmetry can otherwise overstate one model’s advantage
• –The entertainment diagnostic is a useful secondary lens, but Bradley-Terry ranking is the metric that should drive the headline