EVA-Bench - Accuracy: leaderboard

Metric: EVA-A pass@1 (%; share of trials in which task completion is exact, judged faithfulness is at least 0.5 and speech fidelity at least 0.95; 213 scenarios in three enterprise domains, five clean-audio trials each, averaged equally over the airline, healthcare HR and IT service management domains). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.

Top models

#ModelScore
1gpt-realtime-1.546.7
2Gemini 3.1 Flash Live29.2
3Ultravox-Realtime27
4GPT-Realtime-mini16.3

Interactive version: theaggregate.ai/benchmark?slug=eva-bench-accuracy · How It Works · Data refreshed daily, snapshot 2026-09-26.