Dialog-RSN-1 is the fastest model Veris AI has tested

An independent benchmark found Dialog-RSN-1 replies faster than every other dialog model tested.

  • Matt Henderson

    VP Research

  • 16 Nov 2026
  • 2 min
Median reply delay against cost per call, for every agent Veris tested. Source: Veris AI.
Median reply delay against cost per call, for every agent Veris tested. Source: Veris AI.

Share

Independent benchmark VAmoS Pro Bench found Dialog-RSN-1, PolyAI's dialog model, replies faster than anything else tested.

Veris AI, an independent voice-agent testing lab, runs VAmoS Pro Bench by putting an AI agent on an electric utility's billing line and sending it simulated customer calls. Their latest tests show Dialog-RSN-1 replies faster than every other voice agent tested.

Veris's test

Imagine this scenario: you call about your electric bill. You finish your question, but nobody answers. You say "hello?" just as the agent on the other end starts talking, and you end up talking over each other.

Voice agents can struggle with getting the timing right, so Veris measures that pause on every turn of every call.

In Veris's test, Dialog-RSN-1 took the same 300 calls as every other agent on the board, from the same simulated callers. The callers are calm or angry, and they call from a quiet room, a street, a café or a room with the TV on.

Why Veris calls Dialog-RSN-1 "seriously fast"

Half of Dialog-RSN-1's replies start within 1.24 seconds after the customer stops talking. Most other agents take more than two seconds.

The chart below plots reply delay against cost per call for every agent Veris tested.

Scatter chart titled VAmoS Pro: the most realistic voice-agent benchmark, plotting median reply delay against cost per call. PolyAI Dialog-RSN-1 has the fastest median reply delay of every agent tested, on the Pareto frontier alongside LiveKit and Smallest AI.
Median reply delay against cost per call, for every agent Veris tested. Source: Veris AI.

How Dialog-RSN-1 works

Dialog-RSN-1 is audio-native: one model handling turn-taking, hearing, and replying together, instead of passing audio through a separate speech recognizer first. That removes a step that otherwise adds delay and splits hearing from understanding.

Because Dialog-RSN-1 hears the caller and holds the whole conversation, it already knows which names, products, and account numbers to listen for, and the first token it writes decides whether to reply yet. It still reasons when a turn calls for it, typically in under half a second, and we serve it on our own hardware with speculative decoding and our own quantization, so it often writes several tokens at a time.

Get started with Dialog-RSN-1

Dialog-RSN-1 is available now. Learn more here.

Teams evaluating a dialog agent for a live deployment can request a walkthrough of these results and how they map to a specific use case.

Full results for every agent on the board are published by Veris AI on its VAmoS Pro leaderboard. Veris AI's own write-up of this benchmark is here.

Hear a voice agent on your own calls

Book a 30-minute demo and we’ll run a conversation from your world through a live agent.