Does a newer AI model make a better voice receptionist?
Not necessarily. Regal's October 2026 benchmark shows newer models losing on live calls because they dropped fast settings. Here is what that means for the model behind a client's phone line, and why Block's Dessa is still only a demo.
Lumina7 min read
No. A newer model does not make a better voice receptionist, and October's data shows why. Regal's October 2026 benchmark found that models tuned for coding and agent work often score worse on live calls, mostly because they dropped the fast, low-reasoning settings a phone line needs. Before you move a client's line to anything with a bigger number in its name, check how long it takes to start talking.
That is the short version. The long one matters if you sell AI phone systems to plumbers, dentists or anyone else whose customers hang up after three seconds of silence.
What did Regal's October benchmark find?
Regal runs what it calls the Voice CX Benchmark. It grades models turn by turn against real call transcripts, and it sets one hard bar for live voice: time to first token under 1,300 milliseconds. The numbers below come from Regal's October 2026 write-up, which I read in full. Costs there are relative and based on public API pricing, so treat them as direction, not an invoice.
GPT-6 Sol, released September 22, took the top Voice CX score at 786. Its time to first token was 870 ms. Regal puts it about 22% cheaper than GPT-4.1 and about 30% cheaper than GPT-5.6 Sol.
Fine. Now the models that lost.
GPT-6.1 Sol shipped a week later and lands roughly level on quality. Its lowest setting, though, takes 1,711 ms to start, and Regal says reasoning-off is no longer accepted. Same price, nearly double the wait. Gemini 3.8 Flash ranks first at calling a required tool and first at following instructions, then blows the latency bar at 2,516 ms. Claude Sonnet 5.5 costs more than twice what GPT-6 Sol does and has no reasoning-off setting either. Regal passed on all three.
Regal's own summary is blunt. Newer models tuned for coding and agentic work don't necessarily do better in voice.
Why does a slow first word hurt more than a wrong word?
Because the caller can't scroll back.
On a chat widget, a two-second pause looks like typing. On a phone, two seconds of dead air is a dropped call, or a person saying "hello? hello?" over the top of the answer your agent was about to give. Then the agent hears the interruption and restarts. We covered how callers actually judge these systems in does an AI receptionist sound human, and the short answer is that they forgive a flat voice faster than a late one.
Here is the part vendors skip. A reasoning model that "thinks" before it answers is doing exactly what you want in a coding agent. In a call about a burst pipe at 9:40 on a Sunday night, it is a liability. The caller does not want the best possible answer. They want a person-shaped thing to say "I've got you, what's the address?" inside a second.
Which model should a small agency actually pick?
I'd pick the boring one. I know that's an unsatisfying answer.
Regal prioritized two: GPT-6 Sol and DeepSeek V4.1 Flash. The second is the interesting one for margins. Per Regal, it sits 3.2 points behind GPT-6 Sol, starts talking 517 ms sooner, costs about a tenth as much, and ranked first of every model tested on jailbreak resistance. It was reached through Baseten's Model API, which Regal says meets U.S. data residency requirements. It is also weaker on acknowledgment turns and less reliable at calling required tools.
Read that last clause twice. A receptionist that books appointments lives or dies on tool calls. A cheap, fast model that skips the calendar lookup one call in twenty is cheaper only until the first owner finds an empty slot that should have been filled.
Claude Haiku 5.5 is a smaller story. It gained 34.7 Voice CX points over Haiku 4.5, ranks first on intent accuracy and second on context retention, and costs about an eighth of GPT-6 Sol. It also ranks 32nd at avoiding unneeded tool calls and 27th on task progression. Regal is looking at it for intent classification only. That is the right call, and a useful pattern: one model to decide what the caller wants, a different one to hold the conversation.
What about Block's Dessa, the model that runs on a laptop?
Worth watching. Not worth planning around yet.
On October 23, Jack Dorsey announced Dessa from Block's research lab. TechCrunch's coverage calls it a full-duplex speech model, so it listens and talks at once, and says it runs locally on a MacBook M4 Max or better. Block says it was trained entirely on human speech and has no separate language model behind it. In the demos, the agent changed voice and cadence and, in one clip, sounded tired because it said it was tired.
I couldn't find a price, a license, a latency figure, a benchmark, or a release date. Block has not opened public access and has asked safety researchers to test it, especially for cases where a voice agent could deceive people. No API and no number yet means it stays on your reading list and out of your proposals.
It does hint at one thing, though. If speech models get good enough to run on a single workstation, the per-minute model cost that agencies argue about today stops being the interesting line item. We went through that math in what a voice AI agent costs per minute, and the minutes were already the cheap part.
Who should ignore all of this?
Most of you, most weeks.
If you resell a platform where the voice layer is already chosen and tested, model churn is the vendor's problem, which is part of what you pay for. And if your clients are not getting complaints about their line, you have no call to make a change just because a chart moved. I have watched agencies burn a weekend swapping a model on a line that was working, then spend the following week finding out why bookings dipped.
The people who do need this are the ones wiring their own stack from raw APIs. Sometimes that is the right path, and I won't pretend otherwise. If that is you, the comparison in hire a developer versus a white-label stack is worth a read before you sign up for maintaining model swaps forever.
A test you can run in an afternoon
Pull three recorded calls from a client's line. Grab a short call, a rambler, and the ugly one where the caller changed their mind twice. Feed identical audio to the model you run today and to the one you're eyeing. Write down three things:
seconds until the first word comes back
whether the booking tool fired, and fired once
whether the agent ever invented a price
Three calls is no benchmark. If the candidate is plainly worse, though, you'll hear it, and that's all you need to know before live traffic gets near it.
FAQ
Is the newest voice model always the best one for my client's phone line?
No. Regal's October data has newer models scoring below older ones because the newer releases dropped fast settings. A leaderboard rewards quality. A live call needs quality and speed together, and the newer releases traded away the speed.
What latency is too slow for a voice agent?
Regal's bar is a time to first token under 1,300 ms. Gemini 3.8 Flash measured 2,516 ms and GPT-6.1 Sol's lowest setting measured 1,711 ms, and both missed it. Total delay on a real call also includes speech recognition and the phone network, so staying well under that bar is safer than sitting at it.
Should I use a cheaper model to protect my margin?
Only after you have tested tool calls. A model at a tenth of the cost looks great until it skips a required lookup. Run your own recordings first, then decide.
Can Dessa be used for client builds today?
No. Block has not released it publicly, and TechCrunch reports no pricing, licensing or release date. Check back when there is an API.
What to do this week
Open the demo on your own site and time how long the receptionist takes to answer after you stop talking. Count it out loud. If it feels slow to you, it feels slow to a plumber's customer. You can try the same exercise on our live demo, see how the platform handles voice limits, or start an account and send the link to a prospect.
