What an AI voice call actually costs

← All posts

Almost every conversation about AI calling starts with someone guessing the price, and almost everyone guesses high. They are thinking of a call centre seat, or of what a Western voice API charges once telephony is added on top. So it is worth writing the actual figure down.

Three rupees fifty

Sarvam Voice Agents is ₹3.50 per minute, and that rate includes authoring, simulation, phone numbers, evaluation and analytics. That last list matters more than it looks. The phone number is in the price. There is no separate telephony vendor, no per number rental arriving on a different invoice, no carrier relationship to negotiate before you can dial anyone.

If you have priced this stack anywhere else you know how unusual that is. The normal shape is a voice API that bills you for the model, plus a telephony provider that bills you for the call, plus a number rental, plus whatever the two of them do to each other at the boundary when something goes wrong.

What it costs if you assemble it yourself

You can build a voice agent out of parts instead, and Sarvam publishes those rates too: speech to text at ₹30 per hour, text to speech at ₹30 per ten thousand characters, and the 105B model at ₹29.28 per million input tokens against ₹73.20 per million output.

I am deliberately not going to multiply those into a total and present it as the DIY price, because the honest answer is that it depends entirely on how much each side of the call talks, and any number I produced would be an assumption wearing a decimal point. What I will say is that once you have added a telephony provider and a number to the assembled version, the managed one at ₹3.50 stops looking like the expensive option. That surprised me.

What our own demo line has cost

There is a box on our homepage where you type your mobile number and the agent calls you. That runs on our Sarvam account, not yours, so we pay for every one of those calls.

Here is the entire bill so far, read out of the database rather than remembered. Seven completed calls. Eight hundred and five seconds of conversation. The mean call ran a hundred and fifteen seconds, the shortest forty five, the longest a hundred and eighty one.

At ₹3.50 a minute that is about ₹47 in total, and roughly ₹6.70 for an average call.

Seven calls is not a data set and I am not going to pretend otherwise. Three of them were us testing. But the shape of the number is the point: the entire public demo of an AI voice product, running live on the internet for anyone who wants it, has so far cost less than lunch.

The part that actually matters

Svarloop does not resell those minutes. When a customer runs a campaign, the calls go to the customer's own Sarvam account, on the customer's own key, billed to them directly. We never see the invoice and we take nothing on top.

That sounds like leaving money on the table, and in the short term it is. Reselling voice minutes with a margin is a completely normal business and a lot of good companies run on it.

But it changes what we are incentivised to want. We earn exactly the same whether a customer places fifty calls a month or five thousand. So when we design a campaign flow, there is nothing quietly pulling us towards more dials. We do not have a revenue reason to loosen the calling window, or to relax the cap of three attempts per number per day, or to make the opt out slightly harder to trigger. Those limits cost us nothing to enforce, which is the only reliable way to be sure they will still be there in a year.

A product that makes money per call and also promises to call responsibly is asking you to trust that it will keep choosing the less profitable option forever. We would rather not be in that position at all.

What this model costs us

It has a real price and it is paid on the demo.

Because we cannot recover the cost of a stranger's demo call from anywhere, every one of those calls is spend with no path to revenue attached. That is why the demo is rate limited harder than we would like: one call per number ever, a small daily ceiling, and a kill switch. Not because the calls are expensive in absolute terms, but because the honest budget for proving the product to strangers is whatever we are willing to lose.

A company that marked up minutes could afford to be far more generous with its demo, and would then make that back on the customers who converted. We took the other trade. Whether that was the right call is a question the next few hundred calls will answer better than I can.

---

Rates quoted from Sarvam's own announcement and published API pricing, both checked on 16 September 2026. Prices change. Check them yourself before you build a plan on them.

More notes from building this are on the blog, including why we moved calling onto the customer's own PC and eight things Sarvam's docs do not tell you.