EVA-Bench - Accuracy: leaderboard
Voice agents handle simulated phone calls from an audio user simulator in 213 enterprise scenarios (airline customer service, healthcare HR and IT service management), calling tools against a scenario database; EVA-Bench (Bogavelli et al., 2026). EVA-A pass@1 in percent: the share of trials in which the final database state is exactly right, an LLM judge rates faithfulness at least 0.5 and a speech judge rates the spoken key entities at least 0.95, over five clean-audio trials per scenario, averaged equally over the three domains; higher is better.
Metric: EVA-A pass@1 (%; exact task completion, faithfulness >= 0.5 and speech fidelity >= 0.95); five clean-audio trials, 213 scenarios, equal mean over three domains. Source: servicenow.github.io. Saturation forecast: Rough model projection: around 2026. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Cartesia Managed Agents - GPT 5.5 | 87.41 | |
| 2 | Grok Voice 3.0 | 73.72 | |
| 3 | Scribe v2.2 Realtime + GPT-5.4 + Eleven Flash v2 (ElevenAgents) | 67.2 | |
| 4 | GPT Realtime 2.1 | 59.04 | |
| 5 | Grok Voice 1.0 | 58.84 | |
| 6 | Gemini 3.6 Flash + Gemini 3.1 Flash TTS Preview | 56.2 | |
| 7 | Gemini 3.8 Live Extended Thinking (High) | 54.81 | |
| 8 | Universal 3.5 Pro + GPT-5.4 + Sonic 3.5 | 53.15 | |
| 9 | Grok Voice 2.0 | 51.73 | |
| 10 | Apriel Voice 27B + Sonic 3.5 | 51.29 |
Interactive version: theaggregate.ai/benchmark?slug=eva-bench-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.