LLM Debate Benchmark puts Sonnet 4.6 first

This is one of the cleaner "debate as eval" designs out there, because it controls for side bias and scores full argument arcs instead of one-shot answers. It still measures a very specific skill, though: sustained adversarial persuasion on contentious topics.

• –Side swapping is the big methodological win; it keeps the leaderboard from rewarding whichever side the model happened to get.
• –Bradley-Terry over paired outcomes is a better fit than raw judge averages, especially when the judges are themselves LLMs.
• –The frontier is crowded, so Sonnet 4.6's lead reads more like a narrow current edge than a permanent gap.
• –GLM-5 leading the open-weights pack is the practical signal for teams that want strong debate behavior without closed-model dependency.
• –MiMo V2 Pro's 10.4% content-block rate shows that operational fragility can matter as much as rhetorical strength.