# GPT-Live 延迟评测：一致性比速度更重要

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-07-23 08:54
- AIHOT 分数：41
- AIHOT 链接：https://aihot.virxact.com/items/cmrwtu9h502cvrobhf9h6lb2q
- 原文链接：https://x.com/rohanpaul_ai/status/2080094099433193653

## AI 摘要

第三方机构 Agora Media Lab 实测了 OpenAI 未公布的 GPT-Live 端到端延迟。结果显示，其最大提升来自响应一致性：延迟标准差从 489ms 降至 104ms，而中位数响应仅改善 205ms。更低的延迟标准差意味着每次回复的启动时间更可预测，而非偶尔让用户长时间等待。

## 正文

Interesting benchmark reports on GPT-Live's Latency.

Claims that GPT-Live's biggest gain comes from consistency， not raw speed， and thats why it feels smoother.

It's median response improves only 205ms （vs previous advanced voice mode）， while latency standard deviation falls from 489ms to 104ms.

Lower standard deviation here basically means each reply starts predictably， instead of occasionally making you wait much longer.

### 引用推文

> Agora：@OpenAI didn't publish GPT-Live's end-to-end latency. So Agora Media Lab measured it on an iPhone - with repeatable speech, dual-track waveform recordings, and ...