What the audit looks like
One example voice stack, priced three ways: tune what it runs now, move to a cheaper API, or self-host open models. This is the arithmetic a Voice Cost Audit starts from, before anything is tested on your calls.
Example only. Built from public vendor list prices and ILLATE's own arithmetic. Not a client result, and not measured on real calls.
Every number carries a tag: LIST a vendor's published price, ESTIMATE our arithmetic from list prices and the inputs below, ASSUMED an input chosen for this example. Prices checked on 21 Oct 2026 on each vendor's own page.
Want to see the method run end to end? The demo runs open speech-to-text models on 30 synthetic calls and prices three paths on a different assumed stack, with the measured numbers marked.
The example stack
A cascaded voice agent on a hosted platform: the caller's audio goes to speech-to-text, the text to an LLM, and the reply to text-to-speech, over a phone line. Every input here is ASSUMED, a choice for the example; a real audit replaces each one with yours.
| Input | Value |
|---|---|
| Stack | A hosted voice-agent platform (Vapi-style) running Deepgram Nova-3 speech-to-text, GPT-5.4 and ElevenLabs Flash text-to-speech, answering inbound calls on a Twilio US local number |
| Volume | 100,000 connected minutes a month, about 25,000 calls |
| Call length | 4 minutes on average; Twilio's round-up adds half a minute a call on average |
| Silence | 15% of each minute, nobody is speaking |
| LLM turns | 3 a minute, 60 output tokens each |
| System prompt | 4,000 tokens, resent on every turn and not cached, because a per-call detail at its top breaks prefix caching. With 400 other input tokens a turn, 91% of input tokens are the repeated prompt |
| Bot speech | 500 characters a minute |
What one connected minute costs today
The all-in cost per connected minute, split by what you pay for. About $0.123 a minute, or about $12,300 a month at 100,000 minutes.
| Line | $ / min | Share |
|---|---|---|
| Platform fee (Vapi hosting) LIST | $0.0500 | 41% |
| LLM: GPT-5.4, 13,200 input and 180 output tokens ESTIMATE | $0.0357 | 29% |
| of which the repeated, uncached system prompt ESTIMATE | $0.0300 | 24% |
| Speech-to-text: Deepgram Nova-3, every connected second LIST | $0.0077 | 6% |
| Text-to-speech: ElevenLabs Flash, 500 characters ESTIMATE | $0.0200 | 16% |
| Telephony: Twilio inbound, rounded up to the minute ESTIMATE | $0.0096 | 8% |
| All-in cost per connected minute ESTIMATE | $0.1230 | 100% |
Silence has no line of its own. The platform and streaming speech-to-text both bill every connected second, so the 15% of silence is already inside those lines (about $0.0012 a minute of the speech-to-text line). Retell, AssemblyAI streaming, the Deepgram Voice Agent API and OpenAI's GPT-Live also bill session time, so moving to one of them keeps silence on the bill.
Twilio rounds each call up to the next full minute. On 4-minute calls that adds about 12.5% to telephony: $0.0011 a minute, under 1% of the bill.
List prices only: no discounts, volume tiers or promotions. Deepgram is shown at its regular $0.0077 a minute; a limited-time price of $0.0048 runs now. For scale, Vapi's own estimator puts a Deepgram, OpenAI and ElevenLabs stack at $0.082 to $0.129 a minute.
The same minute, priced three ways
Each path is priced on the same inputs. The audit recommends whichever wins once quality is checked, even if that's staying on a vendor API.
Tune the current stack
$0.0960 / min, −22%
ESTIMATE
Keep every vendor. Move the per-call details below the fixed prompt so its 4,000 tokens cache: GPT-5.4's cached input lists at $0.25 per 1M tokens instead of $2.50, so the LLM line falls from $0.0357 to $0.0087 a minute. The prompt itself doesn't change, so the risk to answers is low; latency and answers still get checked.
A cheaper API
$0.0496 / min, −60%
ESTIMATE
Leave the platform for a speech-to-speech API. Mistral's Voice Agents API lists $0.04 a minute with Mistral Small 4, plus the same telephony. Grok Voice 3.0 lists $0.08 a minute, which comes to −27% here. The model and the voice both change, so answers, latency and the voice need checking, and you build the phone bridge the platform gave you.
Self-host open models
$0.0446 / min, −64%
ESTIMATE
Open speech-to-text such as Parakeet or Whisper on CPU servers, open text-to-speech such as Kokoro on a GPU, and your own orchestration, with GPT-5.4 kept by API and path A's prompt fix. Assumed servers: two AWS c7i.4xlarge and two L40S on RunPod Secure Cloud, always on so one of each can fail, at list prices of $0.714 and $1.09 an hour. That's $0.0104 and $0.0159 a minute. Their capacity is assumed, not tested, and engineering and on-call time aren't included.
| Path | $ / min | A month | Change |
|---|---|---|---|
| Before | $0.1230 | $12,300 | – |
| A: tune | $0.0960 | $9,600 | −22% |
| B: cheaper API | $0.0496 | $4,960 | −60% |
| C: self-host | $0.0446 | $4,460 | −64% |
Every row in the table is an ESTIMATE: the inputs above times list prices, at 100,000 minutes a month. On paper B and C are close, and that's where the audit's quality gates decide. A path that fails word error rate, p95 latency or a listening check on your calls is out, whatever it saves. Paths also combine: path A's prompt fix is already inside path C.
B and C don't include any extra cost for the phone bridge, orchestration, monitoring or paid support they would need. In a real audit those charges count toward the 20% test and would shrink both savings; engineering and on-call time sit outside it.
Licences of the open models named: Whisper (MIT / Apache-2.0), NVIDIA's parakeet-tdt-0.6b (CC BY 4.0), Kokoro (Apache-2.0).
How the refund rule would apply
Would this example clear the 20% test?
The rule: the audit fee is refunded in full if the possible savings come to under 20% of your current all-in cost per connected minute, from your last 30 days of invoices.
In this example the best path on paper is C at −64%, and path A alone comes to −22%. If at least one path also passed the agreed quality gates on a client's calls, the savings would clear 20% and the fee would stand; if none passed, the fee would be refunded in full.
This box shows how the rule works on an assumed bill. It doesn't establish savings or refund eligibility for any team: a real audit starts from your invoices, and a path only counts once it passes the quality gates on your calls.
What a real audit adds
This page stops at list prices. The audit replaces every assumption with your numbers and checks quality before it recommends anything.
- Your invoices. Your last 30 days of voice bills, so the starting cost per minute is what you actually paid, discounts and volume tiers included.
- Your calls. A sample of your calls, tested inside your infrastructure; I never take a copy. Your real call length, silence, turns and prompt size replace the inputs above.
- Quality gates. Word error rate on your calls, p95 latency against your target, and a listening check of any new voice. A path that fails a gate is out, whatever it saves.
- One recommendation. A written plan for the path that wins, what each step costs, and the saving to expect.
Price your own stack
The Voice Cost Audit runs this on your invoices and calls: $1,000 fixed, or ₹75,000 for India-based teams, with a full refund if the possible savings come to under 20%.
Where the numbers come from
- Vapi hosting fee and cost estimator (vapi.ai)
- Deepgram Nova-3 streaming and Voice Agent API (deepgram.com)
- OpenAI GPT-5.4 token prices (developers.openai.com)
- OpenAI cost optimization guide (prompt caching, GPT-Live billing) (developers.openai.com)
- ElevenLabs API prices (elevenlabs.io)
- Twilio US voice prices (twilio.com)
- Twilio per-minute rounding (help.twilio.com)
- Mistral Voice Agents API (docs.mistral.ai)
- Grok Voice 3.0 (docs.x.ai)
- Retell billing during silence (retellai.com)
- AssemblyAI streaming billing (assemblyai.com)
- AWS EC2 on-demand prices (aws.amazon.com)
- RunPod L40S (runpod.io)
The arithmetic behind every ESTIMATE is the inputs in section 01 times these list prices. Prices change; the audit uses the rates on the day it runs.