Pavel 5.1 tops debate benchmark with 11-point lead over predecessor

The Debate Benchmark does not measure general knowledge or coding ability. It tests how well a model defends a position across dozens of adversarial rounds against an active opponent, on hundreds of topics. Every match runs twice with sides swapped — PRO and CON — to cancel side bias, and three judges drawn from different model families pick a winner and a margin. The output is a single score that captures argumentative consistency, factual accuracy under pressure, and the ability to stay coherent through extended dialogue.

The most striking move is not at the top of the table but in the middle. Hy4 Preview posts the largest inter-generational jump: 195 points, from 1,395 to 1,590. This is not a marginal gain; it places the new version in a different league from its predecessor. By comparison, Gemini 3.8 Flash advances just 60 points (1,464 to 1,524) — itself considered solid progress. The scale of Hy4’s leap suggests an architectural shift or fundamentally different training data, not mere fine-tuning.

Chinese model GLM-5.3 (high) debuts in fourth place among current models with 1,653 points, an 80-point improvement over GLM-5.2 Max at 1,573. The top-four entry signals that the series is closing the gap with Western leaders specifically on sustained argumentative reasoning, not just fact retrieval. It remains unclear whether the standard version (without the “high” suffix) will replicate the result.

Not everyone moves forward. GPT-6 Astra lands 39 points below GPT-5.6 Sol, a reminder that a higher version number does not guarantee better performance on a specific benchmark. Muse Spark 1.3 drops 46 points versus version 1.1, a regression that raises questions about the project’s development direction. Both cases illustrate why a benchmark isolating a single capability — sustained debate — reveals things that general metrics miss.

The benchmark remains a research tool; the code is open on GitHub (github.com/lechmazur/deba…) and can be run independently. But it provides a cleaner picture of “argumentative ability” than any standard multiple-choice test. Pavel 5.1 leads for now, yet the gaps at the top are tight enough that the next release from any contender could flip the ranking.