Dialog-RSN-1 live latency: turn decision to first model output

For the week of 30 Oct to 6 Nov 2026, the median time from turn decision to first model output on live production voice calls on Dialog-RSN-1 was 164 ms, with a p90 of 409 ms, across 1,407,255 generations.

What this measures

True p50 and p90 over all live production Dialog Studio (formerly Agent Studio) voice calls on Dialog-RSN-1 in US production. The clock starts when the voice agent has decided the caller finished and enters generation. It stops when the first streamed model output of any kind, text or a completed tool call, reaches the voice agent. Most first outputs in this measure are completed tool calls, not spoken words.

This is not the model's request-to-first-token time. On most generations the model started while the caller was still speaking, during turn detection, so its output can already be waiting when the clock starts.

What it leaves out

End-of-speech detection, model work before the turn decision, tool time, text-to-speech, playback and the phone network. This is not what a caller hears.

These figures describe this one week. They are not a guarantee.