AgentJudgeBench (Hard, With Ground Truth) - GPT-5.4 Outputs: leaderboard

Metric: Judge alignment (%; agreement of the judge's per-metric verdicts (tool selection, parameter structure, call sequence, query coverage) with the programmatic reference on the hard (most ambiguous) rewrite of 3,808 BFCL-style tool-calling records over six workflow-DAG topologies; judged outputs from GPT-5.4 at temperature 0; the judge sees the ground-truth tool calls). Source: arxiv.org. Saturation forecast: Around December 2026. 6 models tracked.

Top models

#ModelScore
1GPT-5.481.4
2Claude Sonnet 4.580.7
3GPT-OSS-120B80.7
4QwQ-32B80.5
5GPT-OSS-20B79.4
6Gemini 2.5 Pro (Preview 05-06)77.6

Interactive version: theaggregate.ai/benchmark?slug=agentjudgebench-hard-with-ground-truth-gpt-5-4-outputs · How It Works · Data refreshed daily, snapshot 2026-09-29.