EVA-Bench - Faithfulness: 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). Faithfulness in percent: an LLM judge's 1-3 rating, rescaled to 0-100, of whether the agent's statements and actions stay grounded in its instructions, the domain policy, tool results and what the caller said, over five clean-audio trials per scenario, averaged equally over the three domains; higher is better.
Metric: Faithfulness (%; mean LLM-judge rating mapped from 1-3 to 0-1); 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 | 77.79 | |
| 2 | Flux General Multi + GPT-5.4 + Sonic 3.5 | 75.57 | |
| 3 | Universal 3.5 Pro + GPT-5.4 + Sonic 3.5 | 74.57 | |
| 4 | Grok Voice 3.0 | 66.24 | |
| 5 | Scribe v2.2 Realtime + GPT-5.4 + Eleven Flash v2 (ElevenAgents) | 65.34 | |
| 6 | Grok Voice 1.0 | 58.87 | |
| 7 | Gemini 3.6 Flash + Gemini 3.1 Flash TTS Preview | 56.37 | |
| 8 | Ink 2 + Haiku 4.5 + Sonic 3.5 | 53.58 | |
| 9 | GPT Realtime 2.1 | 49.42 | |
| 10 | Scribe v2.2 Realtime + Gemini 3 Flash + TTS Conversational v3 (ElevenAgents) | 45.72 |
Interactive version: theaggregate.ai/benchmark?slug=eva-bench-faithfulness · How It Works · Data refreshed daily, snapshot 2026-11-12.