For voice-agent vendors
Get your voice agent independently benchmarked
Cozzo helps businesses pick the right AI voice agent with honest, comparable test results. If you’d like your product in our next round, tell us about it below. Results are independent: nobody can pay for a better one.
Latest results
Voice agent latency on real phone calls
Round 2 timed 85 test calls and sessions across 4 platforms (Vapi, Retell AI, ElevenLabs Agents, Ultravox) on a real phone line and over each platform’s browser connection, from Ashburn, Virginia (US East). This is the method a new vendor would go through.
What we test
Phone latency
We call your agent on a real phone number and time the gap between the caller finishing a sentence and your agent starting to answer, on every turn. Reported as median and p90.
Browser latency
The same script over your direct realtime (web) connection, so buyers can see how much the phone line adds.
Transcription accuracy
Word error rate on what the caller said, from our recordings and a fixed reference script.
Talk-over
How often your agent starts speaking while the caller is still talking.
Verified pricing
What the test calls cost on your public plans, checked against your official pricing page and linked so anyone can verify it.
How the method works
Same job for everyone
Every vendor gets the same scripted caller and the same short receptionist task. In Round 2 that was a dental-office booking call in English.
Default settings
We test the setup a typical customer would start with, and we publish exactly which models and settings were used. You can tell us which plan or configuration is recommended for new customers.
Real lines, real recordings
Calls go over a real phone number from a single, disclosed region. Timings come from our own caller-side recordings, not vendor dashboards.
Limits in plain view
Sample sizes, dates, region and every known caveat are published next to the numbers.
The full protocol is on how we test and the methodology page.
Independent, by design
- No vendor can pay for a better result, a better position or a softer write-up.
- Vendors can’t edit, approve or block results before or after they’re published.
- If you find a factual error, send us evidence and we’ll re-check it. Corrections are dated and public.
- Sponsor placements on Cozzo are always labeled and kept separate from benchmarks and rankings.
Cozzo tip: I sing the same song for every vendor. Same script, same line, same stopwatch.
Inquiry
Ask to be included
Tell us about your product. Cozzo’s Relations team reads every inquiry and will reply by email.
Questions
Can we pay to be included or to rank higher?
No. Inclusion is an editorial decision and results are measured, not negotiated. Nothing a vendor pays for changes a benchmark result.
Will we see results before they’re published?
We’ll tell you when results for your product go live and what was tested. You can’t review or change the numbers, but you can report factual errors with evidence at any time.
What do you need from us?
A product URL and, ideally, a contact who can confirm the recommended default setup. Tests run on accounts Cozzo controls, the same way a new customer would sign up.
Which products can be tested?
Voice agent platforms and AI receptionists that can answer a phone call. Speech APIs can be included in transcription tests.
Just want a free directory listing? See get listed.