EVA-Bench - Task Completion: leaderboard

Metric: Task completion (%; share of trials whose final scenario-database state matches the ground truth exactly; 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-1.573.9
2Ultravox-Realtime47.3
3Gemini 3.1 Flash Live47.3
4GPT-Realtime-mini34.5

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