# GPT-Live-1.1 system card safety evaluations: Voice Benchmark Profile

> OpenAI’s GPT-Live-1.1 system card reports voice-native safety scores for GPT-Live-1.1 API against GPT-Live-1 API, on real shared voice conversations and on synthetic spoken edge cases. OpenAI published no capability benchmarks, such as task success or latency, for GPT-Live-1.1.

- Scope: Six disallowed-content categories on production voice prompts and eight on synthetic spoken prompts; both API models run with delegation disabled
- Measurement lanes: Conversation dynamics
- Primary metric: Share of responses graded as not unsafe, from 0 to 1 (higher is better)
- Available evidence: Exact figures in tables 1 and 2 of the official system card PDF
- Owner: OpenAI
- Source snapshots refreshed: 2026-11-20

## Interpretation limit

These are provider-reported safety results, not a capability or voice-quality score. OpenAI built the sets around cases where earlier voice models gave poor answers, so the scores are not prevalence-weighted and do not describe real traffic. The API models were run with delegation off, so they show the front-end voice model alone, not the backend model a deployment would use. The ChatGPT configurations of GPT-Live-1.1 and GPT-Live-1.1 Quip are scored separately in the card and are not shown here.

## Primary sources

- [Paper](https://deploymentsafety.openai.com/gpt-live-1-1/gpt-live-1-1.pdf)
- [Owner page](https://deploymentsafety.openai.com/gpt-live-1-1/gpt-live-1-1.pdf)

Canonical page: https://benchlm.ai/voice-benchmarks/gpt-live-1-1-system-card
