AI Calling Cost in India: What Goes Into a Per-Minute Rate

Almost every AI calling platform quotes a price per minute of conversation. The number is easy to compare and easy to misunderstand, because the per-minute rate is the sum of several separate costs, and the choices you make when configuring an assistant change each of them. This article explains what is inside the rate, what moves it, and how to estimate what a pilot will cost before you run it.

Azmth CMS advertises AI calling from INR 2 per minute; the applicable rate depends on the configuration you choose. The mechanics below apply to any platform built on the same components.

The four cost components

An AI voice call runs a pipeline: audio in, text out, a reply generated, audio back out, all over a phone line. Each stage is a service with its own price.

• Speech to text (transcription): converts the customer's speech into text continuously during the call. Priced per minute of audio. Providers such as Deepgram and Sarvam charge different rates, and Indian-language models can be priced differently from English.
• Language model (LLM): reads the conversation and your system prompt and writes the next reply. Priced per token of input and output. A long system prompt is re-read on every turn, so prompt length has a direct cost.
• Text to speech (voice): speaks the reply. Priced per character or per minute of generated audio. Premium voices from providers such as ElevenLabs cost more than standard voices; Sarvam Bulbul voices for Indian languages sit in their own tier.
• Telephony: the actual phone call, priced per minute by the carrier (Plivo, Twilio, Telnyx) and dependent on destination. Domestic Indian mobile termination and international calling differ substantially.

On top of per-minute usage there is usually a number rental: a monthly charge for the phone number you call from, debited when you buy the number and renewed monthly.

Realtime speech-to-speech models (from OpenAI or Google, for example) collapse the first three stages into one model with a single per-minute price. That price is higher than a pipeline of separate components but the latency is lower.

What changes the rate

Two assistants on the same platform can cost very different amounts per minute.

1. Model choice. A frontier LLM costs more per token than a smaller model from Groq or Cloudflare. For a scripted reminder call, the smaller model is often sufficient.
2. Voice choice. A premium cloned voice can be several times the price of a standard one.
3. Language. Indian-language transcription and synthesis via Sarvam are priced differently from English via Deepgram; check both if your audience is bilingual.
4. Prompt length. Every turn re-sends the system prompt. A 2,000-word prompt on a 3-minute call with 20 turns costs far more in LLM tokens than a 300-word prompt.
5. Call duration. A reminder call that completes in 45 seconds costs a fraction of a qualification call that runs four minutes. Billing granularity matters too: ask whether calls are billed per second or rounded up to the minute.
6. Bring your own keys. If you connect your own OpenAI, ElevenLabs or Deepgram accounts, those components are billed to you by the provider at your negotiated rates and the platform charge covers the rest.

Estimating a pilot

A pilot estimate needs three numbers: how many contacts, what share will answer, and how long an answered call lasts. Unanswered calls still consume telephony for the ringing time, but no AI minutes.

Worked example, purely illustrative:

• 500 contacts, 40% answer rate = 200 conversations.
• Average answered call 2 minutes = 400 conversation minutes.
• At an all-in rate of INR 2 per minute = INR 800 for the conversation component.
• Plus number rental for one Indian number for the month, plus telephony for the 300 unanswered attempts at the carrier's ringing charge, if any.

Run that pilot, then read the actual per-call cost from the call records rather than the estimate. Most teams find the biggest lever is duration: tightening the prompt so the assistant reaches the outcome faster reduces cost more than switching providers.

Comparing platforms fairly

When comparing per-minute quotes from different platforms, ask each vendor the same questions:

• Does the rate include telephony, or is that billed separately by the carrier?
• Is the rate for a specific model and voice, or a blended average?
• Is billing per second or per minute rounded up?
• Are unanswered calls charged?
• What is the number rental and are there setup fees?
• Is there a platform or seat fee on top of usage?
• Can I bring my own provider keys, and what does the platform charge then?

A platform quoting a low per-minute rate that excludes telephony, uses a small model and rounds up to the minute can end up more expensive than one quoting a higher, all-inclusive rate.

Cost per useful outcome

The per-minute rate is an input, not the result. The number that decides whether AI calling pays off is the cost per useful outcome: per confirmed appointment, per qualified lead, per promise-to-pay. Two campaigns with the same per-minute rate can differ five-fold on that number depending on list quality, language selection and prompt design.

Read the transcripts from the first hundred calls. If the assistant is taking too long to reach the question that matters, shorten the path. If customers are hanging up in the first ten seconds, fix the opening line or the language. Those changes reduce both the cost per minute you consume and the cost per outcome you get.

Where to check the current rate

Azmth CMS shows the applicable per-minute rate for your selected providers, voice and model in the dashboard pricing view, and wallet transactions list every top-up and deduction. The public pricing page at cms.azmth.in/pricing describes the advertised starting rate and the components. For a campaign-specific estimate, sales@azmth.in can quote from your expected volume, average duration and languages.