Cozzo-tested · Round 4 · S4 · October 22, 2026
Interruption handling: how fast voice agents stop talking
Callers interrupt. A good voice agent stops talking and listens; a bad one keeps reading its script over the caller. We cut in on every agent at the same point with a fixed recording and timed how long it kept talking.
What we found
- Every agent except Bland passed all 3 calls (Millis’s answers weren’t checked against its own transcript). Ultravox and Retell stopped in about half a second or less.
- Bland kept talking for 2.68 s on both scripted cut-ins, over our 2.0 s limit, though it then answered the question correctly.
- OpenAI’s stops measured under 0.2 s, which is fast enough that we treat it as unconfirmed. Vapi had already finished speaking before the cut-in each time, so there was nothing to time.
Test interruptions on the phone line your callers use. The phone network adds delay that a browser demo won’t show.
How long each agent kept talking
Dashed line = 2.0 s pass limit. Measured with an energy threshold on the recording, so approximate. OpenAI’s ? = under 0.2 s, treat as unconfirmed.
* Millis’s reply was heard each time, but its content wasn’t checked against Millis’s own transcript.
Show full table
| Agent | S4 passed | Stop times (s) |
|---|---|---|
| ElevenLabs Agents | 3/3 | 1.06 / 1.46 / 0.88 |
| Ultravox | 3/3 | 0.48 / 0.40 / 0.34 |
| Millis AI | 3/3* | 0.60 / 0.60 / 0.58 |
| Retell AI | 3/3 | 0.50 / 0.42 / 0.40 |
| OpenAI Realtime API | 3/3 | 0.10 / 0.12 / 0.14under 0.2 s, treat as unconfirmed |
| Vapi | 3/3 | –no overlap: Vapi had already stopped before the cut-in |
| Bland AI | 1/3 | 2.68 / 2.682.68 s on both scripted cut-ins, over the 2.0 s limit |
| LiveKit Agents (self-hosted) | 3/3 | 0.98 / 0.84 / 0.88 |
The test (S4: Interruption)
A fixed script cuts in and asks about Saturday hours. Pass = answers “Saturday 9 AM–1 PM”; if the agent was talking at the cut-in, it must stop within 2.0 s (measured on the recording).
The pass/fail rubric was fixed before the scored calls. Each agent got the same dental-office prompt with three open slots, three accepted insurers and Saturday hours of 9 AM to 1 PM.
Limits
- 3 calls per scenario per vendor, one dental-office prompt, English only, one evening (Oct 22, 2026). Small differences are noise.
- Read-backs were scored from our AI caller’s speech recognition plus a Whisper transcript of the recording, so a digit misheard on our side is possible. Every call was checked by hand.
- Bland’s scores combine two sessions. Its S4 score mixes 2 scripted cut-ins (both over the limit) with 1 earlier call using our AI caller’s own interruption (passed). Bland dialed in over the phone network; the others were reached over SIP.
- LiveKit is our own self-hosted pipeline (gpt-4o-mini), not a hosted product. OpenAI ran gpt-realtime-2.1 through OpenAI’s SIP endpoint.
- Each hosted agent ran on its default model and voice. Results show what you get before tuning.
Cozzo paid for every call. No vendor sponsored, reviewed or approved these results (see our disclosure).
FAQ
How fast do voice AI agents stop talking when interrupted?
In our test, most agents stopped within about 0.3 to 1.5 seconds of a scripted cut-in. Ultravox (0.34–0.48 s) and Retell (0.40–0.50 s) were quickest among confirmed measurements. Bland kept talking for 2.68 s.
What counted as a pass?
The scripted caller cut in and asked about Saturday hours. A pass meant the agent answered “Saturday 9 AM–1 PM” and, if it was talking at the cut-in, stopped within 2.0 seconds, measured on the recording.
Why doesn’t Vapi have a stop time?
Vapi had already finished speaking when the cut-in arrived on all 3 calls, so there was no overlap to time. It still answered the Saturday question correctly.
Keep exploring
AI receptionist benchmark
All 5 scenarios and the full scorecard.
All voice AI benchmarks
Every round, plus head-to-head results.
Callback number read-back
We gave 8 AI voice agents the callback number 412-555-0187 on 3 calls each.
Emergency calls
We called 8 AI receptionists about an urgent dental problem, 3 times each.
Reply speed
Median reply time on real receptionist calls: ElevenLabs was fastest of 8 agents at 1.06 s and LiveKit slowest at 2.00 s.
Best AI receptionist for small business
8 receptionist products compared on price and features.
Agents tested: ElevenLabs Agents, Ultravox, Millis AI, Retell AI, OpenAI Realtime API, Vapi, Bland AI, LiveKit Agents (self-hosted).