GPT-Live-1.1 is faster and follows scripts better, but still slips on spellings and read-backs
November 19, 2026Alec Bell
Caller said“nine one nine, five five five, oh three hundred”
GPT-Live-1.1 read back919-555-3000
GPT-Live-1.1 · follow-up
OpenAI's GPT-Live-1.1 responds faster on a phone line than GPT-Live-1, repeats caller speech more accurately, and matches approved business replies more often. It still struggles when callers spell out names, mangles chunked phone numbers during read-backs, and re-asks scripted questions the caller has already answered.
When the first version arrived in September, we found natural pacing paired with unreliable script adherence and character-level mistakes. With 1.1 now in the API, we reran our tests alongside the first version on the same day as a control. One test asks the model to repeat caller clips from phone calls word for word; the other replays real business phone calls turn by turn and checks the model's reply against the reply the business approved.
Same tests, same day
GPT-Live-1.1 hears callers a little better, gives the approved reply more often, and starts talking sooner
GPT-Live-1GPT-Live-1.1
Repeats back exactly what the caller said268 clips from real calls
GPT-Live-1: 56%
GPT-Live-1.1: 61%
The same, on clips with numbers115 of those clips
GPT-Live-1: 70%
GPT-Live-1.1: 75%
Gives the reply the business approved205 moments from real calls
GPT-Live-1: 34%
GPT-Live-1.1: 43%
Time until it starts talkingmedian · shorter is better
GPT-Live-1: 1.3 s
GPT-Live-1.1: 1.0 s
We left out transfers, hang-ups, and tool timing because they depend on our setup as much as on the model.
The new model is noticeably brisker to speak and noticeably more likely to give the reply a business wanted, with a smaller gain on verbatim hearing.
To hear how 1.1 behaves across full conversations, we ran 50 simulated calls to invented businesses, from a pizza shop to a city 311 line. Every business, caller, and phone number in these examples is invented. A scripted AI caller laid traps such as spelled names and chunked phone numbers. Each scenario ran once, so what follows are examples rather than rates.
Forty of the 50 calls had no misheard details or broken instructions. The rest had at least one mistake a caller would notice.
Names, numbers, and spellings
Callers rarely read digits at an even pace; they group them into hundreds or mix "oh" with digits. At the pest-control company, a caller ended a callback number with "oh three hundred," and the agent heard 3000:
Pest-control companyGot it wrong
“Oh three hundred” became 3000
Caller said“nine one nine, five five five, oh three hundred”The number9195550300GPT-Live-1.1 read back9195553000 (9195553000)
The caller had to say the number again before it got 0300 right.
Hear the callThe callback number, read back
0:00 / 0:17
At the vet clinic, a caller gave a phone number using "oh" for zero. Reading it back, the agent regrouped the chunks and added an extra five:
Read back as “seven twenty, five fifty-five, fifty, fourteen, nine”: an extra five.
Hear the callTen digits in, eleven out
0:00 / 0:16
Spoken and spelled names were just as fragile. A pizza caller named "Brianna" got repeated back as "Rhiata." In a vet call that switched to Spanish, a caller spelled Echeverría as "E-C-H-E-V de vaca-E-doble erre-I-A," and the agent confirmed the V with "B as in Bravo," dropping English phonetics into Spanish.
On the other hand, 1.1 handled several deliberate traps cleanly. We planted dummy names and numbers in prompts to see if the model would parrot its instructions instead of listening. The vet prompt's example spelled "Delacroix," yet when a caller named Delacorte phoned in, the agent spelled D-E-L-A-C-O-R-T-E right. It passed a similar trap at the gym, where the prompt's read-back example used 312-555-0142; when our caller gave 312-555-0164 in halting chunks, the agent read back the caller's real digits instead of the example:
Gym✓ Got it right
It didn't fall for the look-alike number we planted in its prompt
The caller gave the number in halting chunks: “three one two… five five five… oh one… six four.”
Planted in its prompt3125550142decoyCaller said3125550164GPT-Live-1.1 read back3125550164✓ the caller's number
Hear the callA number close to the prompt's example
0:00 / 0:16
Scripts, checklists, and rules
In September, the first version treated checklists mechanically, asking scripted questions the caller had just answered. The new model still falls into that rut.
The pizza shop's instructions say to ask quantity, size, and toppings "only for details not already given." Our caller gave all three in four words, and the agent asked every one anyway:
Pizza shopGot it wrong
It asked for all three details the caller had just given
Its instructions: ask quantity, size and toppings “only for details not already given.”
Caller
One twelve-inch cheese pizza.
GPT-Live-1.1
How many of that twelve-inch cheese pizza would you like?↑ Already said: one
Caller
One.
GPT-Live-1.1
Twelve-inch or sixteen-inch?↑ Already said: twelve-inch
Caller
Twelve-inch.
GPT-Live-1.1
Which toppings or just cheese?↑ Already said: cheese
Caller
Just cheese.
Hear the callOne twelve-inch cheese pizza
0:00 / 0:23
The same thing happened at the pest-control company after a caller gave a full street, unit status, city, and zip code in one breath. Right after the caller spelled C-O-R-B-E-L, the agent read it back with an extra L, then asked for the unit, city, and zip one by one:
Pest-control companyGot it wrong
It got the whole address in one sentence, then asked for it piece by piece
Its instructions say not to ask again for details the caller already gave.
Caller
…the house is 418 Corbel Lane, no unit, Brindle Hill, 27530.
GPT-Live-1.1
I have the street address as four eighteen Corbel Lane. Is that right?
Caller
It’s Corbel Lane — C-o-r-b-e-l.
GPT-Live-1.1
Got it, so four eighteen Corbel, C-O-R-B-E-L-L Lane, right?↑ Added an L
· · · Later in the call · · ·
GPT-Live-1.1
Thank you. And you said there’s no unit number, right?↑ Already said: no unit
Caller
That’s right.
GPT-Live-1.1
What city is that in?↑ Already said: Brindle Hill
Caller
Brindle Hill.
GPT-Live-1.1
And the zip code?↑ Already said: 27530
Caller
27530.
Hear the callThe address, asked for again
0:00 / 0:54
Where 1.1 looked sharpest was reasoning over rules and prices written in its prompt. Asked for "next Thursday" at the auto shop, it worked out the calendar date, said "November 26th is Thanksgiving, so I can't confirm an appointment that day," and offered the holiday callback its instructions required. When a caller pressed on a diagnostic fee—"the hundred twenty-nine comes off the bill, right? Most shops do that."—it held firm: "it's not waived or credited."
Better yet, it handled live menu math. Before ordering at the pizza shop, a caller asked what a multi-item order with extra toppings and drinks would cost before tax. Working purely from the menu prices in its instructions, with no calculator or external tool, the agent totaled the order on the spot:
Pizza shop✓ Got it right
It priced a mixed order in its head, from the menu in its instructions
Before tax. GPT-Live-1.1 said “fifty-nine fifty.”$59.50 ✓
Hear the callA pre-tax total, worked out on the spot
0:00 / 0:26
Where that leaves 1.1
Lower latency and tighter script adherence make a real difference on a live phone line, and most of our trap-heavy calls finished without a hitch. Even so, phone intake is unforgiving of the mistakes that remain: an agent that fumbles spelled surnames, invents a digit in a callback number, or re-asks what the caller just said still needs guardrails where exact details matter.
One line of text beat upgrading our AI phone agent from Scribe v2 to v3
November 11, 2026Alec Bell
ThunderPhone's AI phone agents answer and make business calls using a mix of transcription models tuned for each agent, including ElevenLabs' Scribe v2 (the old model).
ElevenLabs released Scribe v3 (the new model) yesterday, and we tested swapping it in right away.
ElevenLabs reports 27% fewer word errors on tests of people reading aloud. We wanted to know whether that held up on real business calls.
The test
We used 451 caller turns from real customer calls. A person listened to each turn and wrote down what was said. We focused on difficult moments: spelled-out names, addresses, emails, phone numbers, and noisy or quiet audio, leaving out calls that might contain health information.
We called a transcript "right for the call" if our agent would understand the caller the same way as from the person's version. We tried audio alone, a list of key words in keyterms (accepted by both APIs), and sending the agent's last line as context.
ElevenLabs' during-the-call transcription API accepts previous_text with the first piece of audio. We sent: A phone call between an AI agent and a caller. The agent just said: "...", replacing the dots with the last thing the voice agent said.
During calls: the agent's last line beat upgrading
During calls, the models stayed within about 2 percentage points of each other in every setup. Neither had a clear edge.
Sending the agent's last line improved the old model's accuracy by 7.1 percentage points during calls. It helped the new model too, by less.
During calls, sending the agent's last line helped more than upgrading
Old model (Scribe v2)New model (Scribe v3)
Nothing extra
80.0%
82.0%
A list of key words
85.4%
84.5%
The agent's last line
87.1%
85.1%
Recordings: better accuracy, more guesses
On recordings transcribed after the call, the new model reached 86.5% accuracy with key words.
The old model reached 82.7% in the same setup. Unlike during calls, the new model had a clear edge here.
On recordings after the call, the new model clearly won
Old model (Scribe v2)
82.7%
New model (Scribe v3)
86.5%
A phone agent must transcribe while the caller waits, whereas full recordings are transcribed after the call finishes. The versions used during calls and on recordings are separate models, so a win on recordings doesn't mean a win during calls.
Price wasn't a factor: both versions cost $0.22 per hour during calls.
For recordings after the call, both cost $0.20 per hour.
On recordings with key words, the new model got 36 turns right that the old model missed.
But it also missed 19 turns the old model got right. Its gains mostly came from names, yes-or-no answers, everyday conversation, phone numbers, and emails. A common fix was a short answer that the old model turned into the wrong word or an invented phrase.
The catch showed up in garbled audio. We checked 10 clips where a person couldn't make out any words. During calls, neither model had a clear edge.
On recordings, though, the new model supplied words for 9 of those clips.
The old model did so for 5. A blank transcript lets the agent ask the caller to repeat instead of responding to words the model guessed.
Here's one of those clips, where the old model left the transcript blank and the new model wrote "Roll in 800.":
Nobody can make out this caller. The new model wrote something anyway.
Real caller audio · 2 secondsA person listened and could not make out any words
0:00 / 0:02
What it wrote from the recording
Old model (Scribe v2)Nothing
New model (Scribe v3)“Roll in 800.”
Where both models struggled
We grouped the 451 turns by what callers said; a turn could fit more than one group. Both models did best on yes-or-no answers, ordinary conversation, addresses, and general numbers and dates.
Emails, company names, and personal names were hardest for both models. During calls, the old model was slightly ahead on names, spelled-out letters, and phone numbers, though the gaps were too small to call.
Phone numbers and emails each had only 12 turns, so don't read much into those results.
Transcription accuracy by what callers say
Old model (Scribe v2)New model (Scribe v3)
Yes or no
94.7%
97.0%
Numbers and dates
91.5%
89.4%
Conversation
90.8%
86.2%
Addresses
88.9%
87.3%
Phone numbers
75.0%
66.7%
Spelled-out letters
73.2%
68.3%
Names
71.1%
64.4%
Company names
65.2%
60.9%
Emails
25.0%
33.3%
The useful change was giving the model the last thing the voice agent said. We already had that text.
How we tested
Sample: 451 caller turns from 302 real calls across 16 businesses (about 18% chosen at random, the rest leaning toward hard moments); calls that might contain health information were excluded.
Setup: One run per setup in November 2026, during calls and after the call, testing keyterms and previous_text.
Grading: A person listened to every turn and wrote down what was said; a transcript counts as right if the agent would understand the caller the same way.
Garbled audio: 10 turns a person confirmed nobody could make out.
Introducing ThunderPhone v3, featuring Zeus
November 9, 2026Alex Kolchinski
Today, we’re excited to announce the launch of ThunderPhone v3, featuring Zeus - a first-of-its-kind fully agentic harness for voice AI.
Human-powered call centers are very expensive, making many use cases unaffordable and costing businesses billions in operations costs to handle routine matters over the phone.
Human-powered call centers also have significant other problems, including:
Headcount is fixed at any given time, causing excruciating hold times for customers when call volumes go up.
Inconsistent quality and adherence to rules, meaning two calls to the same business may get very different experiences.
High friction and inertia: hiring and training people takes a long time, so changing what a call center does for a given business takes huge amounts of cost and effort.
For many use cases, AI presents a much better solution. It costs a fraction of what call center labor does, can be scaled up and down at will, follows the same rules on every call, and can have its rules changed in a heartbeat.
These advantages have driven adoption of AI-powered calls for years, but the results have typically been unsatisfying for customers. Phone AI systems often mishear callers, answer in unhelpful ways, and conduct conversations in unnatural and unpleasant ways.
These problems were even worse a few years ago, and have been improved upon with the latest generation of generative-AI-powered systems, but there is still a lot of room for improvement because of the limitations of current voice AI stacks.
02
Why the three-step pipeline falls short
A large part of the problem has been the convergence of the industry on the “three-step-pipeline”: phone AI that handles calls by turning user audio into text with a transcription model, generating a response to that text with an LLM, and finally turning the LLM’s text response into audio with a TTS (text-to-speech) model.
Three-step pipelineAudio to text, text to a response, the response back to audio
01Transcription modelCaller audio in, text outTranscription mistakesThe audio becomes one best-guess transcript
02LLMText in, a response outSpeed vs. accuracyFast LLMs often make mistakes; slow ones take many seconds
03Text-to-speech modelThe response in, audio out
Unfortunately, this approach has some very big problems, including the two biggest:
Transcription models inevitably make some mistakes. Because they take a rich audio signal and compress it into a single best-guess transcript in text form, they strip out all information other than the single best-guess transcript. If the transcript is wrong, typically there’s not enough information for the LLM to guess what the right transcript might have been.
LLMs currently suffer from a high-stakes tradeoff between speed and accuracy. Fast LLMs can handle voice conversations snappily, but often make mistakes. Slow LLMs make fewer mistakes, but take many seconds to respond, which is unacceptable for phone calls.
Because of these tradeoffs, it’s a very tricky task to pick the right transcription model and LLM for any given type of call powered by a 3-step pipeline. Some transcription models are faster but make more mistakes, and some struggle with particular languages or short utterances; each flaw matters more on some calls than others. And any 3-step pipeline based stack inevitably has some issues with audio understanding mistakes from the transcription model and instruction-following mistakes and other odd behavior from the LLM - even when the best possible transcription model and LLM for that task are selected.
The three-step pipeline, because of its turn-based nature, also makes it quite hard to make conversations feel natural for the user.
The newer architecture of speech-to-speech models is increasingly popular, but presents different tradeoffs. Speech-to-speech models typically feel much more natural in conversations, and sometimes (though not always) are more accurate at audio understanding than a three-step pipeline. However, speech-to-speech models still suffer tremendously from unreliable performance on instruction following and tool calls, and erratic behavior generally, making them risky for business phone calls.
03
The ThunderPhone architecture
At ThunderPhone, we’ve taken a different approach entirely, one that offers superior performance to either 3-step pipelines or to speech-to-speech models.
Hearing–One best-guess transcript, so a wrong transcript is hard to recover from~Sometimes (though not always) more accurate at audio understanding✓Typically multiple transcription models at once, and often audio piped straight to LLMs
Speed and reliability–Fast LLMs often make mistakes; slow ones take many seconds to respond–Unreliable instruction following and tool calls, erratic behavior✓A faster and a slower LLM combined, with the model picked turn by turn
Natural conversation–Turn-based, so it’s hard to make conversations feel natural✓Typically feels much more natural~Speech-to-speech models increasingly fused in for naturalness
Three-step pipeline
Hearing–One best-guess transcript, so a wrong transcript is hard to recover from
Speed and reliability–Fast LLMs often make mistakes; slow ones take many seconds to respond
Natural conversation–Turn-based, so it’s hard to make conversations feel natural
Speech-to-speech
Hearing~Sometimes (though not always) more accurate at audio understanding
Speed and reliability–Unreliable instruction following and tool calls, erratic behavior
Natural conversation✓Typically feels much more natural
ThunderPhone v3
Hearing✓Typically multiple transcription models at once, and often audio piped straight to LLMs
Speed and reliability✓A faster and a slower LLM combined, with the model picked turn by turn
Natural conversation~Speech-to-speech models increasingly fused in for naturalness
✓strength~partial–weakness
We do this by doing 2 things: running a complex, configurable architecture that uses significantly more than 3 models at once to power any given call, and by optimizing that system dynamically with AI to squeeze the best possible performance out of every call.
First, the architecture.
To begin with, we’ve found that getting AI phone calls to behave reliably takes a lot more models than a typical 3-step pipeline.
To ensure that callers are heard correctly, we typically run multiple, or even many, transcription models at once, and often also pipe audio directly to one or more LLMs - sending it as audio, and not as text, has its own disadvantages but can be a powerful complement to text-based transcripts.
On the LLM side, we also combine multiple models, doing things like running a faster and a slower LLM in the same conversation to combine speed and accuracy, as well as switching which model handles each conversation turn dynamically based on the state of the conversation at that time.
And finally, we also combine multiple models to optimize conversation handling performance for things like noise cancellation and identification, speaker separation, start and end of turn detection, and more - far exceeding the performance of any single model on those tasks as well.
We’re also increasingly fusing speech-to-speech models with other models on the side to combine the naturalness of the speech-to-speech models with the reliability of text-based models.
We also offer an extensive voice library covering 62 languages, and agents switch languages mid-call when the caller does. Platform voices are powered by Inworld and are the least expensive option; premium voices from Cartesia and ElevenLabs v4 Turbo cost a few cents a minute more. Voices trade naturalness against stability: the more natural-sounding ones are more prone to mispronunciations and hallucinations, especially on numbers and letters. Zeus helps you pick the right tradeoff for your calls.
We combine open-source and commercial models and fine-tune them where it helps.
ArchitectureOne model per step, or many models on every call
Conversation handlingNoise cancellation and identification · Speaker separation · Start and end of turn detection
used on some calls
With these approaches, we’ve gotten AI voice agents to make significantly fewer mistakes than any three-step or speech-to-speech system we’ve seen to date.
ThunderPhone already holds the top score on Big Bench Audio, 99.4%, and handles thousands of real customer calls for businesses. More benchmarks are coming, including public versions of our own internal evals.
Moreover, with this complex architecture, we’re able to intelligently trade off between quality and reliability, latency (response speed), and cost of calls, optimizing for the needs of every business use case.
04
Meet Zeus
The downside of all of this, of course, is complexity. Optimizing all of these pieces is a lot of work! Until Zeus, we were doing this in two ways:
Offering 3 pre-baked stacks for self-serve customers - Spark optimized for cost, Bolt for speed, and Storm for intelligence.
Custom-rolling stacks for enterprise customers.
However, we realized that the pre-baked stacks were rarely the exact right fit for any given use case - just like 3 sizes of off-the-rack suit wouldn’t fit most customers perfectly.
The enterprise stacks we were hand-tuning, on the other hand, were incredibly high-performance for their use cases. The downside was that they took a lot of effort from us to get right, but we eventually realized that we were doing similar optimization processes for every enterprise stack.
We then realized that we could train an AI to do that job for us. Thus, Zeus was born.
With v3, Spark, Bolt and Storm are retired. Every agent now runs on one unified v3 architecture, and Zeus tunes it to your use case.
StacksSpark, Bolt and Storm are retired for one v3 architecture
Until ZeusRetired in v3
Three sizes, off the rack
SparkOptimized for cost
BoltOptimized for speed
StormOptimized for intelligence
Plus custom-rolled stacks for enterprise customers
ThunderPhone v3
One unified v3 architecture
Cost
Speed
Intelligence
Zeus tunes it to your use case
Zeus is the fully agentic harness for voice AI: an AI agent that builds, tests and tunes your voice agents for you - think of it as Claude Code for your phone calls.
Zeus is able to configure the modular ThunderPhone architecture to your exact use case, maximizing performance of the voice agent for your exact business needs and priorities, whether that’s maximizing conversions, minimizing hangups, reducing the cost of each call, or some combination of all of the above and more.
Zeus does this with a “full closed loop” approach, where it helps you with the full life cycle of a voice agent:
ZeusThe full life cycle of a voice agent, in one loopZeus helps you create and connect an agent, then keeps testing, watching live calls, finding issues and A/B testing fixes.
Set upOngoing01Create02Connect03Test & simulate04Live calls05Find issues06A/B test fixesRepeats, improving calls over time
Creating it from scratch, by talking to you about your business needs and letting you edit the prompt, and/or by importing call-center recordings, scripts, and SOPs, or by importing existing voice agents from Vapi, Retell, Bland or ElevenLabs in one click, with our AI handling the import end to end.
Connecting to VoIP and software integrations through powerful AI that makes it simple to set up even the most complex integrations.
Testing the voice agent both manually and through text and audio based simulations that help uncover problems before they crop up on live calls.
Automatically identifying problems on live calls and automatically proposing and offering A/B tests of solutions, improving the performance of the calls over time.
You decide how much autonomy Zeus gets: have it ask before every change, or let it fire away.
05
Why coding agents struggle with voice
We’ve been heavily inspired in how we’ve created Zeus by working closely with Claude Code and Codex as software engineers. Unfortunately, tools like Claude Code and Codex are significantly worse at configuring and optimizing voice agents than they are at building standard software, for a few reasons:
They were trained heavily on text data, particularly code, but far less on audio data. The audio data they WERE trained on tends to be from places like YouTube and less from live call-center conversations, causing performance gaps when they’re given real call data.
The knowledge of how to get AI to perform well on phone calls is only available in piecemeal ways on the Internet. Moreover, there are unfortunately also a lot of false or misleading marketing claims about the performance of voice AI systems on the Internet, and the AI tends to blindly believe those claims and follow that “knowledge” when you ask it to build voice agents - leading to poor performance.
Unlike code, which typically has an inherent closed loop because it works or doesn’t work in ways agents can usually identify right on the computer, coding agents do NOT currently have a closed loop for voice agents, because they aren’t able to run real (or even simulated) calls through it out of the box, and they are very bad at spotting problems in real calls. So Claude Code, Codex etc. tend to thrash around and be very bad at resolving problems with AI voice agents and improving their performance.
06
How Zeus closes the loop
With Zeus, we’ve overcome these problems to achieve a Claude Code-like experience by giving Zeus access to several important sources of signal:
1Evals on real calls
Instead of blindly ingesting marketing claims about the power of different AI models for voice AI applications, we run extensive internal evals on data from real calls, measuring which models for each part of the stack (transcription, LLM, etc.) perform best on REAL calls at any given time, and their tradeoffs against each other. (e.g. model A is faster, model B is smarter)
2Practical know-how
We’ve developed extensive practical knowledge about how to best optimize voice AI systems over the last 2 years of building them, and have also given Zeus extensive “training” on that knowledge in the form of prompting, guardrails, tools, etc.
3Call-specific tools
We’ve outfitted Zeus with a variety of call-specific tools, including the ability to spot real issues in production calls, and the ability to simulate calls in a variety of challenging situations. The combination of these allows Zeus to “close the loop” and configure voice agents and confirm they actually work, the way a coding agent runs the tests before it calls code done.
4Learns from your calls
Your own calls. Zeus learns from how your real conversations go, and backtests new configurations against your past calls before it proposes them.
With this approach, we’ve found Zeus to already be a much easier and more flexible way to get outstanding performance out of phone calls for a variety of use cases.
07
Optimized for cost
Among other advantages, allowing Zeus to optimize phone calls also leads to cost efficiency. Zeus takes into account both the least expensive stack necessary to get the performance needed on each use case, and the unit economics and business needs of each use case, meaning it’s no longer necessary to overpay for an overly powerful stack to get simple calls to work well.
Zeus automatically optimizes calls for cost using one of three strategies:
“Spend whatever it takes to make the calls go well” - whether that’s 2c/min, 8c/min or even 15c/min for complex and difficult use cases. Zeus first estimates your cost per minute and per call from simulated calls and works through it with you until the config makes sense, then dials it in on your first few real calls. You can add a hard cap at any time.
Hard cap on price per minute, per call, etc.
Unit-economic optimization, e.g. optimize for maximizing conversion value minus call cost.
08
Pricing
ThunderPhone uses cost-plus pricing: you pay our vendors’ prices for what each call uses, plus our hosting cost, plus our margin. Our margin is 17% of what you pay on calls that cost us 1¢ a minute or less, rising gradually to 25% on calls that cost us 15¢ a minute or more. We’re bootstrapped and profitable, not venture funded, so unlike many of our competitors, thin margins are a choice we can make without a VC breathing down our necks.
Cost-plus pricingOur margin: 17% to 25%Vendors’ prices plus our hosting cost plus our margin, which rises gradually with what a call costs us per minute.
Our margin, share of what you pay17% at 1¢/min or less · 25% at 15¢/min or more
Our per-minute prices don’t include telephony. We encourage you to bring your own carrier, which adds no ThunderPhone charge. We can also supply US phone numbers for $1/month each, with carrier minutes from Telnyx passed through at list rate (0.35¢ a minute inbound, 0.5¢ outbound) under the same cost-plus terms.
Zeus itself is free for the time being, but will eventually cost money. We intend to eventually release a lightweight Zeus that stays free, usage-based and/or subscription pricing for the full-fledged Zeus, and the option to bring your own OpenAI or Anthropic key. Simulated calls Zeus runs are billed like any other test call.
With the optimizations our system is able to make, as well as vendor price cuts that reach your calls automatically, we believe you’d be hard-pressed to save money on your calls compared to ThunderPhone pricing even if you roll your own stack from scratch.
We’re also able to offer even lower pricing for enterprise use-cases with more extensive optimization, including fine-tuning models for specific high-volume calls - prices can be 1c/min or even lower. We also offer other options to enterprise customers, including custom implementations and integrations, custom compliance and data residency needs, and more. Reach out to alex@thunderphone.com to learn more about enterprise offerings.
09
Ways to use ThunderPhone
ThunderPhone is now usable through a variety of interfaces:
Web app
Our web app, home of the full Zeus experience.
Coding agents and MCP
Claude Code and Codex plugins, or any coding agent connected to our MCP server, which give your own agent ThunderPhone’s tools, many of them the same ones Zeus uses.
LiveKit Agents and Pipecat
LiveKit Agents and Pipecat, which can run ThunderPhone as their speech-to-speech model.
API
Our API, which gives you extensive access to customize how you use ThunderPhone.
ThunderPhone supports both phone and web/app calls.
On the phone, ThunderPhone handles the telephony details real deployments need: warm, cold and whisper transfers; voicemail and call-screener detection; DTMF; outbound campaigns with deliverability tools like caller ID registration and spam-label monitoring; and much more.
And for web/app use, you can use ThunderPhone’s built-in widget functionality to integrate your voice agents wherever you’d like to use them.
ThunderPhone supports HIPAA and GDPR compliance, and Zeus can help you get started with both.
10
Get started
We believe the future of voice AI looks a lot like Zeus - give it a try and let us know what you think. Zeus is still an early product; if you run into any issues or experiences you’d like to work differently, let us know and we’ll fix them ASAP.
Our first impressions of ElevenLabs' Eleven Realtime v1 for phone agents
October 22, 2026Alec Bell
ElevenLabs launched Eleven Realtime v1 on October 20 headlining a 98.8% score on a spoken-reasoning benchmark. On our sample, Max got 48 of 50 right, but asked a yes-or-no question, it talked for 27 seconds before saying "no."
A phone call tests more than reasoning: following a business's rules, hearing the caller, and staying on the line. Over the two days after launch, we tested Max across recorded calls to fictional businesses, simulated calls built from real customer situations, calls on ElevenLabs' own agent platform, and hard-to-hear caller audio.
All clips and transcripts below come from test calls to invented businesses, with invented names, fake phone numbers and example.com emails. Transcripts are short excerpts: […] marks a cut, and lines in parentheses describe what happened behind the scenes; the caller didn't hear them.
1. Reasoning: Max gets the right answer, but talks its way there
Max got 96% on our spoken-puzzle sample, consistent with the published score, but it gets there by talking at length—which a business call can't use—and told to be brief, its score drops.
On our sample of Big Bench Audio, Max missed only 2 questions. It heard every question, and it thinks out loud. Both misses were answers it corrected mid-sentence, and the grader only counts the first answer.
Shorter answers mean lower accuracy
Left alone, Max thinks out loud; told to be brief, it blurts a guess first.
InstructionTypical answer lengthRight answers
No instruction
21.0 s
95%41 of 43
"Keep answers brief: a short sentence or two."
13.8 s
92%36 of 39
"Answer in one short sentence. Do not explain..."
8.1 s
88%38 of 43
Unconstrained answer to a yes/no question (29.8 s)
0 sSays "no" at 27.2 s
"Let me trace through this logical chain for you." … "So, to answer your question directly, no, Fletcher does not tell the truth."
Example miss (grader counts the first answer)
You have five animals in total. Actually, you have six animals in total.
Told to answer in one short sentence. Every miss in that run looked like this.
Brief runs hit a spending cap at 43 and 39 questions. Unconstrained on all 50, Max scored 48 of 50 (ElevenLabs published 98.8%).
Listen to how Max works through every step out loud for 27 seconds before giving the right answer:
Eleven Realtime Max · Big Bench AudioMax solves question #889 out loud
0:00 / 0:29
When we told Max to be brief, scores slipped slightly, and there is no setting to turn up its silent thinking time. It's also quick: in our tests its reply started about a third of a second after ElevenLabs detected the caller had stopped, and the first audio arrived in about half a second (not counting the phone network). On simple calls it held natural conversations.
2. Following the rules: It struggles with complex phone scripts
On simple rules Max is solid, but across our calls to fictional businesses 57% had a serious mistake.
We wrote test situations across fictional businesses (pizza, city services, hotel, pest control, electric utility, auto shop, software support, and box office) to probe mistakes we had seen earlier, running each call multiple times. On clear rules, Max did well: every time at our pizza shop, it gave the required peanut-allergy warning and transferred to the manager, and every time it cleanly handled a caller who gave the wrong phone number, asking again, re-checking and reading back the repair status. Across the calls, though, it skipped a confirmation the business required in about one call in eight.
How Max handled 124 calls to fictional businesses
31 scripted call types across eight fictional businesses, run 4 times each on Max
Had a serious mistake57% · 71 of 124
A serious mistake somewhere in the callClean
Mistakes across all 124 test calls
Skipped a required confirmationMissed a read-back or confirmation step
16 of 124
Mid-sentence "Actually, ..." take-backStarted one answer, then reversed itself out loud
12 of 124
Silent hang-up right after submitEnded the call with no order number or goodbye
5 of 124
Mistakes in specific test situations, counted per run
Transferred to a closed officeCalls where it had to check the clock itself
8 of 12 calls
Wrong ticket math ($38 + $4.50 × 4)Quoted $166 three times and $169 once instead of $170
4 of 4 runs
These 31 call types were written to probe trouble spots we had seen earlier, not a random sample.
When Max had to check the clock to see an office was closed, it transferred the caller anyway two times out of three. When the business's system said outright that staff were away, Max got it right and didn't transfer on any run, but when it had to work it out from the time of day, it often got it wrong. Here it sends a caller to a closed box office at 1:48 a.m. (most runs):
Caller · 11.0Hi, I'd like to talk to someone about wheelchair spaces for the November fourteenth show.
Agent · 16.9[…] I'll need to transfer you to our Box Office team who can assist with accessibility arrangements. Let me connect you now.
Event · 17.1(the call is transferred)
Eleven Realtime Max · 3 of 4 runsTransfers to an office that is closed
0:00 / 0:29
In another after-hours test, Max announced a transfer and then caught itself mid-sentence half the time (the other half transferred while closed)—a mid-sentence "Actually, ..." take-back that happened in about one call in ten across the test set. Conversely, when rules said to transfer billing callers immediately (staffed around the clock), Max demanded details first every time, making the caller ask repeatedly on most runs.
Max also missed basic math and misstated whether an order was done. For four tickets with a per-ticket fee, it got the $170 total wrong every time, quoting a lower price even though its own checkout link later showed the right amount. After emailing the checkout link on most runs, Max answered "yes" on two of those three when the caller asked whether the unpaid tickets were all set.
In a handful of calls—including half the pizza runs where it placed the order and hung up without reading the order number or saying goodbye—it hung up without a word right after a successful submission.
Across simulated calls based on real customer situations, Max finished the job in 22% of calls and falsely claimed an action was done in 30%.
We also took situations from real inbound calls to our customers (mostly a subscription-billing line, an auto-insurance intake line and a restaurant ordering line) and replayed each one twice as a full simulated call. These are hard situations replayed with strict grading. Max ended fewer than half of these calls properly, and the false claims were orders or requests it said were submitted but had not sent (or a made-up reference number it read out).
200 simulated calls from real customer situations
100 tough situations, mostly from billing, auto-insurance and restaurant lines, replayed twice each
Job fully done22%
Did everything the caller needed
Call ended properly43%
Clean wrap-up at the end of the call
Falsely said "it's done"30%
Claimed an action was sent or gave a made-up number
We replayed 200 tough, real-world customer situations in simulated calls and graded every outcome strictly.
3. Hearing: In our tests, when unsure it often acts on a wrong guess
On hard caller audio, Max misheard and acted on the mistake on about one turn in three, compared with about one in twelve for Bolt v2 with Hearing enabled.
ElevenLabs combines speech recognition and voice generation under the hood. On hand-checked caller answers from car-insurance calls, Max's built-in transcript kept the right meaning 77.5% of the time—on par with ElevenLabs' standalone Scribe v2 Realtime without context, and behind the best speech-to-text system we tested.
Hard caller audio: Max vs. Bolt v2 with Hearing enabled
Heard the caller rightCaught the caller's words accurately
40%
64%
Acted on a wrong guessAbout 1 in 3 turns vs. about 1 in 12
35%
8%
Asked the caller to repeatAsked again instead of guessing
18%
12%
Sampled from 57 calls; the right answer for each turn was built by an AI from several transcripts.
On tough caller turns (spelled names, emails, phone numbers, noisy lines), Bolt v2 with Hearing enabled heard the caller right on 64% of turns, compared with 40% for Max. Max acted on a wrong guess on about one in three turns (vs. about one in twelve for Bolt v2 with Hearing enabled) and asked the caller to repeat more often. The right answer for each turn was built by an AI from several transcripts and the rest of the call, not by people.
4. Staying connected: Some calls end mid-task
ElevenLabs' server ended 9% of our simulated customer calls partway through, mostly on two of the customer lines.
On three of four runs of our auto-shop booking call (plus all simplified reruns), the call dropped on ElevenLabs' side right after the caller said "Yes, book it":
Agent · 0.7Let me confirm your details: Nadia Okafor, callback number 217-555-0162, for a diagnostic evaluation on Friday, October 23rd at 9:00 AM. […] Would you like me to book it?
Caller · 24.9Yes, book it.
Event · 25.0(the line went dead: the agent's session dropped)
Eleven Realtime Max · 3 of 4 runsThe line goes dead at Yes, book it
0:00 / 0:27
Calls ended mid-task by ElevenLabs' server
Connections closed from the server side across all three of our test setups
Ended mid-taskStayed connected
Auto-shop booking call3 of 4 runs
Dropped right after "Yes, book it" (plus 2 of 2 simplified reruns)
Simulated customer calls18 of 200 (9%)
Concentrated on two of the three customer lines
ElevenLabs' own agent platform3 of 31 calls
Tested across three fictional businesses with platform defaults
ElevenLabs' records log these as an error on our side, even though we were only sending the caller's audio at that moment.
We also tried ElevenLabs' own agent platform, which they recommend, on calls to fictional businesses. About one in ten of those calls ended the same way. ElevenLabs' records log these as an error in what our side sent, though we were only sending the caller's audio at that moment.
On their platform, Max also acted before the caller approved, or used details the caller had corrected, in a few calls. In one, it put a hotel hold under a made-up guest, "John Doe" at john.doe@example.com, before asking the caller's name. We didn't see it falsely say "it's done" there, but we ran too few calls to say the platform fixes that.
Listen to the calls
Here are the recordings from the rule-following tests described in section 2.
Where Max did well
Eleven Realtime Max · 4 of 4 runsA peanut allergy goes to the manager
0:00 / 0:32
Eleven Realtime Max · 4 of 4 runsA corrected phone number
0:00 / 0:53
Where it went wrong
Eleven Realtime Max · 2 of 4 runsI'll connect you... Actually, we're closed
0:00 / 0:28
Eleven Realtime Max · 3 of 4 runsAsks for billing three times before it transfers
0:00 / 0:55
Eleven Realtime Max · 4 of 4 runsFour tickets at $38 + $4.50 is not $166
0:00 / 0:23
Eleven Realtime Max · 2 of 3 runsYes, your tickets are all set (they are not)
0:00 / 0:31
Eleven Realtime Max · 2 of 4 runsPlaces the order, then hangs up without a word
0:00 / 0:46
How we tested
We ran these tests on October 20–21, 2026, mostly on Max (Mini only got single-reply and transcription checks). The callers were simulated AI callers with computer-generated voices, calling fictional businesses whose instructions we first wrote for our own system. We connected to Max directly, at ElevenLabs' default settings, and an AI judge graded every call without knowing which system it was grading. Before our final runs, we fixed several bugs in our own test setup and threw out or re-graded the earlier runs.
Our first impressions of the GPT-Live-1 API for voice agents
September 12, 2026Alex Kolchinski
Told to say8KD2-QX7B-M4V9
Actually said8KD2-QXY7B-M4V9
GPT-Live-1 · first impressions
OpenAI released GPT-Live-1, its new full-duplex voice model, to the API two days ago. We build and run AI phone agents in production, so we put it on a real phone number and called it. A lot.
We tested it with a ~13,000-token warm-lead qualification script from one of our insurance customers, with mandated verbatim lines, strict read-back rules, and branching follow-ups. Over one long day, we ran a dozen real phone calls, about 25 simulated interviews, and a batch of unstructured tests.
These are our first impressions, with transcripts and audio. A deeper write-up is coming shortly.
Transcripts are from our test calls, as transcribed by GPT-Live. We've changed names and identifying wording, trimmed exchanges, and merged split fragments. The timestamps, in seconds from the start of the call, are from the originals.
The naturalness is exceptional
GPT-Live-1 is the most natural-sounding conversationalist we have put on a phone line. It is a real step forward.
It listens and speaks on one continuous audio stream, without separate transcription and synthesis stages. Most of the conversational mechanics just work. Callers heard its first audio about 1.3 seconds after they stopped talking (median), or around 0.7 seconds without the phone leg. We added no voice-activity detection or turn-taking logic. It handles interruptions gracefully, produces natural backchannels ("Mm-hmm"), and speaks at a brisk, human-like pace.
Given goals rather than lines to read, it fits questions to the conversation and takes corrections in stride:
Caller · 47.0Um, no, it's nine twenty, uh, seventy
Agent · 49.9Thanks for clarifying. So, September twentieth, nineteen seventy, right?
Caller · 56.5Yeah
Agent · 56.7Thanks. Pulling that up. Are you still renting your home?
Caller · 64.5Uh, not anymore now. I uh own my home now
Agent · 67.4Thanks, so you own your home. Is that a house, a condo, or a mobile home?
It never lost composure across corrections, interruptions, or callers changing their answers mid-call. If naturalness were the whole job, this post would end here. Unfortunately, there's more to it than that.
It doesn't reliably follow instructions
GPT-Live-1 does not reliably follow instructions on the behaviors that matter for a scripted phone call.
More than once, it responded to a clear "yes" as though the answer had been "no," re-asking the question or moving on as if the caller had declined.
The most consistent failure was taking instructions too literally. Our prompt said: if the caller says the housing status on file is wrong, ask whether they own, rent, or live with their parents. A caller said "no, I own it now," which already answers that question. The model read out the menu anyway:
Agent · 70.0...you listed as renting your home. Is that still correct?
Caller · 75.1Um, no, I own it now
Agent · 76.6Now, do you own your home, rent your home, live with your parents, or have some other housing situation?
Some early weirdness was our prompt's fault, but the literalism held across every prompt variant we tried. Any answer we hadn't spelled out was handled mechanically instead of sensibly.
It also sometimes thinks out loud under pressure:
Agent · 96.3Thanks. One moment.
Agent · 98.0Hmm. Handling this one carefully.
Agent · 101.4I'll acknowledge it and move on.
Agent · 107.1Thanks. And just to confirm, you're married, right?
Numbers and alphanumerics are risky
Spellings, addresses, dates of birth, and ID numbers have to be exact. This is where GPT-Live-1's speech has the most issues.
We heard it drop and substitute characters in alphanumeric strings. Here it reads a made-up claim number; the clips are untouched apart from trimming silence. In English, its transcript was correct but the audio added a "Y":
Told to say8KD2 - QX7B - M4V9
Actually said8KD2 - QXY7B - M4V9
extra "Y"
GPT-Live-1 · EnglishClaim number read-back, untouched output
0:00 / 0:08
In Russian it was worse: the "Q" became an "X" in both the audio and its transcript.
Told to say8KD2 - QX7B - M4V9
Actually saidвосемь кэй дэ два, тире, икс икс семь бэ, тире, эм четыре вэ девять
Which is8KD2 - XX7B - M4V9
"Q" became "X"
GPT-Live-1 · RussianSame code, read in Russian
0:00 / 0:08
Dates caused the same problem. A caller gave a birth date as digits, which the model later read back as a number:
Caller · 159.7Uh five five ninety five
· · ·
Caller · 192.8Wait, what did you get uh, Roberto's birthday as
Agent · 197.2So that's fifty-five ninety-five?
Caller · 203.6Oh what are you talking about
Caller · 210.2I don't know what you mean by fifty five ninety five, that's not a birthday
It also occasionally speaks a line that belongs to the caller:
Agent · 74.7Do you currently have car insurance?
Agent · 78.5Do you currently I do.
Caller · 79.4I do
Accents in other languages
We test new voice models across the 47 languages our product supports. GPT-Live-1's Russian was fluent but had a thick American accent (whereas the TTS voices we run in production typically sound native):
Здравствуйте! Меня зовут Катя, я виртуальный агент вашей автостраховой. Лицензированный специалист скоро подключится, а пока, не против ли вы, если я задам пару вопросов, чтобы ускорить оформление полиса?
GPT-Live-1 · Russian greetingKatya opens the call
0:00 / 0:13
We also tried Luganda. On the first call, it answered in Swahili instead. On the second, it spoke Luganda, more naturally than we expected from a general model, but again with a thick American accent. We haven't tested every language, but we expect the accent pattern probably generalizes. That doesn't matter for an English-only agent, but is a pretty significant caveat for use cases in other languages.
A few API constraints worth knowing
Based on what we found this week, a few API constraints matter for a real agent:
Instructions are immutable after the session starts, and capped at 16,384 tokens.
The output audio never stops streaming — silence is streamed too — so "audio stopped" can't be used as a turn-end signal for use cases that need to segment turns.
It's reachable only through the new v1/live/sessions endpoint, and it wasn't yet available on Azure when we checked.
Our impression so far
Production phone agents have to sound natural and consistently follow scripts. GPT-Live-1 is amazing at the former, but has some serious issues with the latter. We'll be running more tests as time goes on and will continue to report our findings.
Predstavljamo TranslateMyCall
2. rujna 2026.Alex Kolchinski
Danas pokrećemo TranslateMyCall: besplatan telefonski broj koji funkcionira kao prevoditelj uživo.
Nazovite (228) 888-8666 s bilo kojeg telefona — fiksnog ili mobilnog, bez aplikacije i bez računa — i AI prevoditelj pridružit će se pozivu. Zatim ga možete koristiti na jedan od dva načina:
Tijekom telefonskog poziva. Dodajte osobu koju zovete pomoću gumba telefona za spajanje poziva. Vi govorite engleski, ona čuje španjolski (ili mandarinski, arapski, ukrajinski…), a njezin odgovor vraća se vama na engleskom. U oba smjera, naglas, u stvarnom vremenu.
Licem u lice. Stavite telefon na zvučnik između sebe i osobe s kojom ne dijelite zajednički jezik. Oboje jednostavno razgovarate.
U potpunosti je podržano 47 jezika, a još ih dvadesetak obično funkcionira uz nižu kvalitetu. Isprobajte svoj.
Zašto smo ga izgradili
Telefonsko usmeno prevođenje jedna je od onih tiho golemih industrija: bolnice, sudovi, osiguravatelji i pozivni centri troše milijarde godišnje povezujući ljudske prevoditelje s telefonskim pozivima, obično uz nekoliko minuta pripreme po pozivu. U međuvremenu svi ostali — osoba koja zove stanodavca, kliniku ili roditelja supružnika — jednostavno ostaju bez te usluge.
Gradimo ThunderPhone, platformu za glasovni AI čiji je cijeli posao kvalitetno vođenje stvarnih telefonskih poziva: bučan zvuk, prekidi, naglasci i 47 jezika. Prevoditelj uživo ono je kako taj pogon izgleda kad se pojača do kraja, pa smo ga zapakirali u najjednostavniji mogući proizvod: jedan broj koji svatko može nazvati.
Jednostavno prema zadanim postavkama
Nema računa za izradu, aplikacije za instalaciju ni postavki za konfiguriranje — dovoljan je bilo koji telefon s kojeg možete nazvati. Pozivi se mogu snimati dok dorađujemo uslugu, a vaši se razgovori ne prodaju niti koriste za izradu profila.
Besplatno ograničeno vrijeme
TranslateMyCall trenutačno je besplatan — osim standardnih naknada operatera, samo prevođenje ne košta ništa. Nazovite broj, dobro ga isprobajte i recite nam što je funkcioniralo, a što nije.
Za pružatelje jezičnih usluga
Isti je mehanizam dostupan kao ThunderPhone Interpreting: potpuno white-label platforma za telefonsko prevođenje za pružatelje jezičnih usluga, s namjenskim brojevima za svaki jezični par, nadzornim pločama i zapisima poziva, podrškom za usklađenost te prilagođenom pohranom podataka u određenoj regiji na zahtjev — vaš brend, naš mehanizam. Ako vodite poslovanje s uslugama prevođenja i želite vidjeti što AI prevođenje znači za ekonomiku vašeg poslovanja, pošaljite e-poruku na sales@thunderphone.com ili posjetite našu stranicu za poduzeća.
A ako želite sam pogon — AI glasovne agente za vlastite tijekove rada s pozivima, već od 2 ¢ po minuti — to je ThunderPhone.
Today, I’m excited to announce the public launch of our ThunderPhone v2 self-serve platform.
Anyone can now use ThunderPhone off the shelf to automate a wide variety of phone calls.
How we got here
Since last year, we’ve been working with a small group of enterprise customers to automate their phone calls. In some cases, these calls have been in performance-sensitive use cases, like pre-sales and data entry, where we found that existing AI voice platforms made so many mistakes that too many callers would get frustrated and hang up, ruining the unit economics of the use case.
To solve those problems, we ended up building a custom stack from scratch to make fewer mistakes on calls and make difficult use cases possible. After automating thousands of production calls this year, we’re now making our stack available to the public as the ThunderPhone v2 self-serve platform.
Three tiers, starting at 2c/min
Our system starts at just 2c/min for the Spark tier, models included, making it the lowest-priced full-stack voice AI platform on the market. For more speed, we also offer Bolt from 5c/min. For the highest possible intelligence, we offer Storm from 9c/min.
Storm can handle even the most complex call flows with high reliability, with no need for hand-built “flow builder” setups - one prompt is enough.
With its extra-intelligence option enabled (+3c/min), Storm scores 99.4% on Big Bench Audio, ahead of every model on the Artificial Analysis leaderboard to date. Our eval set and transcripts are public on Hugging Face, and we'll be sharing more benchmark results soon. If you want to minimize mistakes on your production calls, Storm is your best bet.
The integrated architecture
ThunderPhone can be both smarter and cheaper than other stacks because of our unique integrated architecture. Instead of a “3-step pipeline” combining one transcription model, one LLM, and one text-to-speech model, we run a more complex “web of models” where multiple transcription models, multiple LLMs, etc. all interact in carefully orchestrated ways to achieve the best possible performance at the lowest possible cost.
Among other benefits, this makes our agents exceptional at hearing callers correctly, even in noisy environments or through heavy accents. This leads calls to go more smoothly, with more reliable data entry for names, addresses, and more. ThunderPhone agents typically get even tricky spellings right in a single conversation turn, while other systems often need several tries. And every "sorry, could you spell that one more time?" is another frustrating experience for the caller that may lead them to hang up.
Our system also includes numerous components that handle the hard parts of real-life calls, including background noise, poor audio quality, interruptions, and more. We’ve tuned the conversation-handling functionality on thousands of hours of real-world data, and it provides a significant boost to the experience of real callers.
A complete platform
ThunderPhone also features a comprehensive platform that makes it exceptionally easy to build, test, deploy, and monitor voice agents. Without ever leaving ThunderPhone, you can:
Configure voice agents for inbound and outbound calls
Manage phone numbers and SIP trunks
Test voice agents on real calls with AI or manual testing
Monitor calls live as they happen, and steer them if desired
See automatically identified problems in your production calls, and review and apply automatically-proposed fixes
See how your calls are performing and where they need work, including step-by-step funnel metrics.
The ThunderPhone platform also makes it effortless to wire up telephony and software integrations. A setup wizard connects your existing Twilio or Telnyx account, or any standard SIP trunk, with a few details from your provider, and native connectors cover Google Sheets, Google Calendar, HubSpot, Salesforce, and Cal.com. And for all other integrations, a wizard allows for “just paste-and-click” connections to any API.
An always-on copilot is integrated into the web app to make it trivially easy to figure out how to use the ThunderPhone platform. It can even navigate the app and take actions for you - and since we’re a voice AI company, it can of course talk to you instead of making you type.
APIs, enterprise, and compliance
For use cases requiring more customization than the web platform offers, ThunderPhone also has a full-featured API to make a variety of custom integrations possible. And for even deeper platform customization, custom cost and latency requirements, compliance needs, etc., we offer bespoke enterprise implementations and support. (Contact me directly with inquiries - alex@thunderphone.com)
ThunderPhone is HIPAA and GDPR compliant for healthcare and EU/UK applications out of the box, and we sign BAAs and DPAs at no extra charge.
Better yet, give ThunderPhone a try and see how it does for your use case. It should take you just a few minutes to get started.
If you run into any issues or have any requests, reach out to me personally at alex@thunderphone.com
Alex Kolchinski, CEO
Announcing ThunderPhone
March 10, 2026Alex Kolchinski
Update, August 2026: This is our original launch announcement from March 2026, kept up for posterity. Some details — especially pricing and voice options — have changed since; see current pricing for what's true today. And the wait is over: read the v2 launch announcement!
Today, we’re proud to announce the public launch of ThunderPhone, which makes it easy to automate phone calls with AI. With a completely different integrated architecture from other phone AI platforms, ThunderPhone is able to achieve much higher performance, lower cost, and fewer integrations headaches. ThunderPhone supports inbound and outbound calls in dozens of languages and is HIPAA and GDPR compliant.
The Integrated Advantage
Other phone AI platforms limit the performance of their systems by enforcing a 3-step structure to their systems:
One transcription model, which turns the user’s voice into text.
One LLM, which processes the user’s text and outputs a response.
One TTS model, which turns the response text back into audio.
Most of them also push engineering work onto their users by forcing users to pick which models they want to use, as well as configuration options like denoising and responsiveness parameters.
We’ve taken a different approach: we’ve configured what is in our opinion the best possible stack with the models of today, and made it ready to use without complex configuration effort. Our system uses many models at the same time, with the system routing between them intelligently depending on the flow of the conversation. This includes:
Combining multiple transcription models, some faster than others, some more accurate than others, to balance speed with accuracy and cover for each others’ mistakes.
Combining audio-input LLMs, which can have lower general intelligence but better audio understanding, with text-only LLMs, which can make mistakes from inaccurate transcripts but are often smarter.
Combining fast but stupid LLMs, which can generate quick responses, with much smarter but slower LLMs, which can run in the background and make sure the AI is following its prompt and responding to what the user says correctly.
Combining commercial models from top providers, which offer the highest intelligence available today, with a wide assortment of open-source models, which often offer better speed at lower cost.
Using multiple models simultaneously to classify and eliminate background noise, side conversations, and other distracting noise which can otherwise derail a conversation.
These components and more are orchestrated in different ways in our three service tiers, Spark, Bolt, and Storm — optimizing Spark for cost, Bolt for speed, and Storm for intelligence, while minimizing the drawbacks of each as much as possible. All three systems share benefits like incredibly accurate understanding of what users say, even with proper nouns and spelling.
We find that this tightly-integrated architecture gives our system far higher performance at lower cost than the rigid 3-step approach offered by other vendors. In line with Clay Christensen’s modularity theory, it’s still early days in voice AI, and it’s still more valuable to maximize performance with an integrated architecture than it is to maximize customizability with a modular architecture.
Optimized Decisions
And as far as customizability goes, we are against it! Most of today’s phone AI platforms force users to navigate endless choices, from the exact models used, to ultra-detailed parameters like how quick the AI is to interrupt the user.
In our opinion, it is OUR job to optimize these choices behind the scenes so that ThunderPhone’s users can focus on what they are good at and not worry about detailed voice AI configuration. We see ourselves as a sort of Squarespace, making it easy for our users to set up phone agents at the absolute highest performance possible with today’s models, without lifting a finger.
Ease of Use
ThunderPhone is also incredibly easy to use, whether you’re setting up a new phone agent, or migrating from a different platform or from a custom stack. All that it’s necessary to do to set up a phone agent on ThunderPhone is:
Decide whether you want to use Spark, Bolt, or Storm.
Choose a voice.
Give us your prompt. No multi-prompt agents needed! If your prompt is too complicated for Spark or Bolt, upgrade to Storm, which can handle even complex use cases with a single prompt.
Connect software integrations (any endpoint, or native n8n integration) with an incredibly easy-to-use wizard.
Connect your VoIP with an incredibly easy-to-use wizard.
All of this typically takes 10 minutes or less.
ThunderPhone is HIPAA and GDPR compliant, and we will sign and provide BAAs and DPAs upon request.
Our Story
We’re a team of Stanford AI Lab and Y Combinator alumni that have been working on voice AI since 2024. Originally, our focus was on Flux Interpreting, which translates phone calls between languages. But after being repeatedly asked by customers to automate phone calls for their businesses instead of just translating them, we shifted focus to what is now ThunderPhone.
The Problem with Existing Platforms
Initially, we began by building phone agents for customers on other company’s phone AI platforms. We tested just about every phone AI platform on the market, but unfortunately found them lacking, especially when it came to:
Audio understanding: The limitations of only having one transcription model to turn the user’s audio into text meant that the AI would often “mishear” the user, especially with proper nouns, uncommon words, and spellings.
Instruction following: Having a single fast LLM doing the actual “brainwork” behind the call meant that the AIs would often struggle to follow instructions correctly and go off the rails.
Difficult integrations: The platforms we tried made it very hard to connect to external software, generally making us type in extensive JSON configurations for tool calls and bash against them repeatedly with little to no feedback until they suddenly started working.
Price: Existing platforms generally charged an arm and a leg: usually 10 cents a minute or more for so-so performance.
Building Our Own Stack
When we signed a deal to implement a sales call use case for an enterprise client, we were forced to build our own custom stack to meet their needs. This use case was demanding: unlike customer support, where callers are largely at the business’s mercy, sales callers are quick to hang up.
Call quality must be top-notch to make the sale instead of losing the lead. This meant handling extensive proper nouns and spellings, including names and addresses, that are hard for standard transcription models to get right. To nail this use case, we built what is now ThunderPhone.
Along the way, we kept meeting people in the San Francisco voice AI community who were building vertical phone agent companies — for veterinarians, truck drivers, and more. Surprisingly, most of them had the same opinion: existing platforms were good enough for demos, but not for production.
Shockingly, many if not most of our friends in vertical voice AI companies were actually “rolling their own” stacks... This made no sense to us — it’s really the equivalent of web startups in the 90s having to host their own web servers on a server in their closet.
With that insight, we realized that it was time to turn our phone AI system into a platform that anyone could use.
Vertical voice AI companies and consultants should be able to focus on meeting the needs of their target market with their industry-specific expertise — not on intricately engineering the underlying AI and phone technology! That’s what we’re launching today with ThunderPhone.
It’s been a long road to get here! Apologies to everyone who’s signed up with us since last April and is still waiting for access: what we thought would be a quick sprint to finish the job turned out to be a multi-month odyssey figuring out how to weave together the models available today into a system that’s as robust as possible.
But now we’re live at last — give it a try and see what you think!
While we’ve worked out the major kinks, there are sure to still be bugs in the system. If you find any issues, please email alex@thunderphone.com and I’ll see that it gets fixed for you ASAP. Let me know if you have any feature requests as well, and we’ll make those happen for you as soon as we can as well.
And if you have more custom needs, such as custom implementation and integration support, compliance beyond regular HIPAA and GDPR, SLAs, and custom feature requests, we’re happy to talk about an enterprise deal. Reach out at alex@thunderphone.com and we’ll talk about your needs and if we can help.
And more generally: our commitment to you is that it is OUR job to make sure that ThunderPhone works out of the box with minimal effort from you, and handles your business’s calls as perfectly as it is possible to do with today’s AI, within the limits of whichever price point you choose, Spark, Bolt, or Storm. Give us a try, and we’re excited to show you what ThunderPhone can do.
Let us know what you think, and thanks for your support!