Performance

off

GPT Live 1

Start a conversation

First audio

—

First action

—

Tool RTT

—

Visible action

—

Barge-in

—
Interruptions only

Latest measured value · trends: last 10 turns

History by model

Connecting storage…

Measurement details

Checking WebMCP…

Same local audio observer for every model. Speech end is estimated using local VAD; audio start observes playback. Visible action follows the committed page paint. Blanks mean unavailable. Tool RTT shows the last call; the export includes every call. Text and clicks use submission time, labelled in history. Overlapping actions can have negative latency.

GPT Live retains its GPT-5.6 Terra backend; Gemini uses native live tools. Each uses the same commerce action handler.

Retention is a model-blind automated assessment after the turn, with manual override. It is an estimate. Transcripts and tool data are saved; microphone audio is not stored.

You can now ask

Tell us what
you’re looking for.