April 22, 2025
Several people have asked why voice agent platform costs seem to be anchoring around $0.10/minute, if the per-minute cost of of using APIs directly and hosting code yourself is $0.02-$0.04/minute ...
Part of the answer is that "hosting code yourself" has more moving parts than just paying for vCPU runtime cost. (The cost calculator linked in the thread below only includes the vCPU cost, benchmarked as AWS on-demand instance cost.)
You'll also need at least:
- service discovery
- load balancing
- logging
- monitoring
- bandwidth
At scale, you'll need:
- more of all of the above :-)
- compliance
- multiple regions
- analytics
- customer support
A voice agent platform has to pay the hard costs for all of these, plus the salaries of the people doing devops.
There are some ways to push costs down below $0.02/minute, though today these require making quality compromises, operating at very large scale with an experienced devops team, or having a specific use case that's different from what most people are doing in the voice AI space.
In general, it's very hard to run an LLM cheaper than Google will serve you Gemini Flash tokens! It's also hard to run a speech model that's cheaper than Cartesia or Deepgram, once you factor in provisioned GPU capacity.
This whole space is evolving rapidly, though. There's some cost pressure on providers, as many commercial and open source models continue to get better and better.
Voice agent cost calculator (and a couple of common calculation mistakes)
Here's a spreadsheet to calculate the per-minute cost of a voice agent.
I cleaned this up to share because I've had the same surprising conversation a few times lately ... 🧵
