vCX-Hard - Non-Hallucination: leaderboard
Real contact-centre calls replayed against each model by Retell AI: the share of turns free of hallucinated facts, in the production configuration with reasoning off. Higher is better. Source: retellai.com/blog/vcx-hard-benchmark.
Metric: Non-Hallucination Rate (%). Source: www.retellai.com. Saturation forecast: Around February 2027. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 84.8 | #33 |
| 2 | GPT-5.4 | 83 | #83 |
| 3 | Gemini 3.1 Pro (Preview) | 82.6 | #64 |
| 4 | GPT-5.2 | 81.3 | #115 |
| 5 | Claude Sonnet 4.6 | 81.2 | #94 |
| 6 | GLM-5.1 | 79.3 | #109 |
| 7 | GPT-5.1 | 75.7 | #145 |
| 8 | Gemini 3.1 Flash Lite | 74.4 | #200 |
| 9 | Gemini 3 Flash | 73.8 | #103 |
| 10 | GPT-4o | 72.8 | #357 |
| 11 | Claude Haiku 4.5 | 72.4 | #287 |
| 12 | Gemini 2.5 Pro | 72.3 | #159 |
| 13 | Gemini 3.5 Flash | 72 | #66 |
| 14 | Claude Sonnet 4.5 | 71.5 | #148 |
| 15 | GPT-4.1 | 71.5 | #258 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vcx-hard-non-hallucination · How It Works · Data refreshed daily, snapshot 2026-11-05.