Smarter Routing Can Slash Voice AI Token Costs (Full Transcript)

A speaker shares benchmark findings showing how duration-aware model routing can lower voice AI costs while preserving accuracy.

Download Transcript (DOCX)

Speakers

Add new speaker

[00:00:00] Speaker 1: So I spent $10,300 last week. Did you miss me? Did you miss me? It's okay if my CFO got mad today, but the truth is smartest engineers in the world do not know how to save tokens. But now you do because we ran an experiment with our benchmark and evals by doing 10,000 plus calls 15 minutes each and guess how much was the cost difference. The $10,000 I spent could have been $100. A $339 bill on Cloud Sonic could have been just $7 on Gemma 431B. Things really change when the duration of calls are longer. If a call is one hour long, you cannot use smaller models anymore. Accuracy is way better with GPT-6 Sol or Cloud models like Sonnet and Opus. So that's why we are helping you by open sourcing our evals. We call it voice standard benchmark and over there, you can just give it to your routing engine on Cloud. It will automatically figure out which model to use saving you 10x the cost in tokens. You can test out yourself by giving hours of audio or seconds of audio which model is best for you and use our LLM gateway for cheaper model cost as well as routing it directly from your voice outputs. Go check out.

Summary

The speaker says their team spent $10,300 in one week on AI voice-call workloads and claims that better model routing could have reduced the bill to about $100. Based on more than 10,000 benchmark and evaluation calls of roughly 15 minutes, they cite an example where a $339 Cloud Sonic cost could fall to $7 using Gemma 3 1B. They stress that call duration changes the optimal model choice: smaller models may be economical for shorter calls, while hour-long calls require more capable models such as GPT-6 Sol, Claude Sonnet, or Opus for accuracy. The team is open-sourcing a Voice Standard Benchmark and offering an LLM gateway to help routing engines select models automatically, reduce token costs, and route voice outputs.

Copy

Download

Title

Cutting Voice AI Token Costs With Smarter Model Routing

Copy

Download

Keywords

voice AI

Remove

token costs Remove

Remove

model routing Remove

Remove

benchmarks Remove

Remove

evaluations Remove

Remove

LLM gateway Remove

Remove

Cloud Sonic Remove

Remove

Gemma 3 1B Remove

Remove

Claude Sonnet Remove

Remove

Claude Opus Remove

Remove

long-duration calls Remove

Remove

cost optimization Remove

Remove

Copy

Download

Key Takeaways

The speaker claims $10,300 in weekly AI-call costs could potentially have been reduced to roughly $100 through better routing.
Their comparison suggests a $339 Cloud Sonic workload could cost about $7 on Gemma 3 1B under suitable conditions.
Model choice should account for call duration, since longer calls may need higher-capability models to preserve accuracy.
The open-source Voice Standard Benchmark is intended to evaluate model performance across short and long audio inputs.
An LLM gateway and automatic routing are positioned as tools for reducing token spend and selecting appropriate models.

Copy

Download

Sentiments

Positive: The tone is energetic and promotional, emphasizing dramatic savings, practical experimentation, and a solution designed to help users optimize voice AI costs.

Copy

Download

Enter your query

{{ secondsToHumanTime(time) }}

Back

Forward

{{ Math.round(speed * 100) / 100 }}x

{{ secondsToHumanTime(duration) }}

Select Audio file