AI Voice Agent Latency Benchmark: What to Measure in Production

Guide section 01

Latency is an end-to-end caller experience

A single model timing does not describe what a customer experiences on a phone call. Measure from the point the caller finishes a turn to the point the next audible response begins. That interval can include endpoint detection, telephony transport, transcription, orchestration, model generation, text-to-speech startup, network delivery and playback.

ReachFly evaluates V4 Voice performance at the conversation boundary because that is where delay becomes noticeable. When publishing a comparison, document the same carrier conditions, geography, prompt complexity, voice, interruption settings and sample size for every platform.

Guide section 02

Publish p50 and p95, not one best-case number

A median response time describes a typical turn, while p95 exposes the slower tail that can make a conversation feel inconsistent. A credible benchmark should report both, explain how many turns were measured and separate answered calls from setup failures.

Do not turn one laboratory result into a permanent competitor claim. Retest after provider, model, routing or telephony changes.

Guide section 03

Measure interruption and turn-taking quality too

Fast speech that talks over the caller is not high quality. Include barge-in response, false interruption rate, end-of-turn detection, long-pause behavior and recovery after crosstalk. These factors often matter as much as raw milliseconds.

Guide section 04

How to compare ReachFlyAI with Retell, Bland or Vapi

Use the same phone network, destination region, prompt, voice style, test script and measurement method. Record the full results and publish the date. If ReachFly measures faster under that controlled test, state the measured result and methodology rather than presenting an unsupported universal claim.