Why We Sell Credit, Not Minutes — and Why Nobody Can Honestly Sell You Minutes

Every voice AI pricing page in India sells minutes. Ours used to. In August we took the minute bundles off and replaced them with call credit, and this post is the working — because the reason is not a positioning preference. It is that we could not make the arithmetic honest.

A voice agent call is not one product. It is a speech-to-text model, a language model, a text-to-speech model (or a single speech-to-speech model doing all three), plus telephony carriage. Each is metered separately, by the vendor, on its own unit — audio seconds, tokens, characters.

So the cost of a minute depends on which models the call ran on. Not slightly. Here are our own three voice bundles, all doing the same job:

• Everyday — ₹4.91/min. The full Sarvam stack — best on Indian languages, and the cheapest to run.
• Natural — ₹9.95/min. Speech-to-speech on Gemini Live. Replies the instant you stop talking.
• Premium — ₹26.37/min. Speech-to-speech on OpenAI realtime. The most capable, and much the dearest.

A customer on Premium burns credit more than 5 times faster than one on Everyday. If we sold both a “500 minute” bundle, one of those two customers would be badly wrong about what they bought — and we would have known which one when we sold it.

On top of the model choice, Indian regional languages — Tamil, Telugu, Kannada, Malayalam, Bengali — run on a dearer speech stack than Hindi or English. Same agent, same script, same call length; different cost per minute.

Which produces the specific failure that made us change: a Tamil clinic and a Delhi D2C brand buy the same plan on the same day, and the clinic runs out first. Nothing is broken, nobody was overcharged, and the clinic is right to feel misled — because the page told them they had bought a number of minutes, and minutes were never the thing being sold.

This is worth knowing before you compare pricing pages, because the minute figures are not comparable and are rarely what they look like.

Look closely at a platform advertising, say, 5,000 minutes at $350, 20,000 at $1,200, 50,000 at $2,500. Divide each price by its minutes and you get $0.070, $0.060, $0.050 — which will be exactly that vendor’s published platform fee, to the cent, on all three tiers. The “included minutes” are the plan price divided by the fee. They are fee-minutes.

Speech, the language model, the voice and the telephony are still charged on top, drawn from a separate balance. So a plan that reads as 5,000 minutes of calling is 5,000 minutes of one line item, and the rest arrives separately. The arithmetic is not hidden — it is just not the arithmetic the number implies.

A plan grants rupees. Starter at ₹2,999/month includes ₹2,500 of call credit, an Indian number and the telephony. Each call draws that credit down by what it actually cost — the pulsed platform fee, the models it ran on, the carriage — and when the credit runs out you top up. There is no overage bill and no invoice at the end of the month, because you can only spend credit you have added.

Our pricing page still shows minutes, but as a range: what the credit buys on the dearest voice, and on the cheapest. The range is wide because the truth is wide. A single number in the middle would be more persuasive and wrong for almost everybody.

Four questions that make per-minute claims comparable:

• Which models does that rate assume? A rate without a model named is not a rate.
• Is telephony inside it or billed separately? Indian carriage is small per minute and not zero.
• What does a regional-language minute cost? If the answer is the same as Hindi, ask what speech model serves Tamil.
• What happens when the bundle runs out? Overage billed in arrears is a different product from credit you top up, and only one of them can surprise you.

Credit is harder to sell than minutes. “₹2,500 of calling” requires a second sentence where “500 minutes” does not, and we lost a clean headline number by doing this. We think the trade is right: a number that is easy to compare and wrong for half the people reading it is worse for us than one that needs explaining, because the explaining happens either way — before the sale, or in month three.