Worked example: one voice minute, priced three ways. See the numbers →
What does one minute of your voice agent cost?
I price three ways to bring it down on your own calls: tune the stack you run, move to a cheaper API, or self-host open models. Then I recommend whichever passes your quality bar at the lowest cost.
Fine-tuned OpenAI model stopping on 23 October? Check it and see what replaces it →
- Three paths, priced on your calls: tune your current stack, a cheaper API, or self-hosted open models
- Voice Cost Audit: $1,000 fixed, or ₹75,000 for India-based teams
- Full refund if the audit finds possible savings under 20% of your current all-in cost per connected minute, from your last 30 days of invoices.
Voice AI cost audit for teams running live calls
For founders and operators paying for speech, model, platform and phone minutes: I measure the cost of one connected minute on your own invoices and calls, then price three ways to lower it: tune your current stack, switch to a cheaper API, or self-host an open model. I recommend the lowest-cost option that passes the quality and latency checks we agree before testing. One component, one language; 10 business days; $1,000 fixed, or ₹75,000 for India-based teams.
Full refund if the audit finds possible savings under 20% of your current all-in cost per connected minute, from your last 30 days of invoices.
Four steps, ten business days
Scope
Scope one costly voice component and one language.
Baseline
Read 30 days of invoices and measure a representative call sample inside your systems.
Three paths
Price the three paths on the same workload and compare quality, latency, reliability and cost.
Recommendation
Receive the evidence table, recommendation and refund decision within 10 business days.
Where one voice minute goes
An example stack priced from public list prices, a hosted platform running Deepgram, GPT-5.4, ElevenLabs Flash and Twilio at 100,000 minutes a month, comes to about $0.123 a connected minute. Your split will differ, which is why the audit starts from your invoices.
| Line | Share of the minute |
|---|---|
| Platform fee | 41% |
| LLM, most of it one system prompt resent uncached on every turn | 29% |
| Text-to-speech | 16% |
| Telephony | 8% |
| Speech-to-text | 6% |
Example only: list prices and our arithmetic, not a client result. The full worked example, with every number tagged →
Three ways to bring it down
On your own calls, I price all three and recommend whichever passes your quality bar at the lowest cost, even if that's staying on your vendor API.
Tune what you run
Keep your vendors and change the settings: prompt caching, voice choice, silence and idle timeouts. In this list-price example, making the system prompt cacheable alone cuts about 22% of the bill; answers and latency still get checked on your calls.
A cheaper API
A different speech-to-text, text-to-speech or speech-to-speech API, priced from list prices and your call profile. The model or the voice changes, so answers, latency and the voice get checked.
Self-host open models
Open speech-to-text models such as Whisper or Parakeet on your own servers. In this STT-only self-host option, caller audio is transcribed on your infrastructure; the LLM and TTS still use APIs. Whether it pays depends on volume, peak concurrency, quality and operating costs. In the demo, self-hosting speech-to-text on CPU did not pay at the assumed 100,000 minutes a month.
Poojith Devan, founder and inference engineer
I run the audit myself. Your call recordings stay in your systems. If an API test needs live audio, you choose the vendor and run it from your own account. The report shows the baseline, each option's assumptions and the checks each option passed. Before ILLATE: ternary LLM quantization and GPU kernels at OneBit AI, and fixed-point inference at Trebuchet Solutions. How the audit works →
Separate from the audit: my OpenAI fine-tune migration work. See the fine-tune research and benchmark →
Get a 20-minute cost estimate
Email me your minutes a month, your languages, and which speech-to-text, text-to-speech and LLM you use. No data needed. I reply within one business day with times for a call.