The Cheapest AI Calling at Scale: How ₹4/min (~$0.05) Actually Works

If you have shopped for an AI calling platform recently, you have seen the price spread: anywhere from $0.05 to $0.50 per minute for what sounds like the same thing — an AI that talks to your customers on the phone. At 50,000 minutes a month, that spread is the difference between ₹2 lakh and ₹20 lakh a year. This article explains where the money actually goes in an AI call, why per-minute prices vary 10×, what volume genuinely changes, and how we price Nutaan at ₹6/min retail, dropping to ₹4/min from just 1,000 minutes and scaling down to just ₹3/min (~$0.036) at volume — among the lowest rates in the market — without cutting the parts that make calls work.

We run this infrastructure ourselves: every claim about cost structure below comes from operating Nutaan's own voice stack in production, not from vendor brochures.

What you are actually paying for in an AI call minute

An AI phone call is a pipeline of four metered services running in real time. First, telephony: the actual phone leg (SIP/PSTN), which in India costs roughly ₹0.30–0.60 per minute wholesale and more on international routes. Second, speech-to-text (ASR): converting the caller's voice into text continuously, billed per audio second by providers. Third, the language model: every conversational turn is an LLM call, with cost depending on which model and how much context (your business knowledge) is loaded into each turn. Fourth, text-to-speech (TTS): the AI's voice, which is the most variable cost of all — premium neural voices can cost as much as the other three components combined.

On top of the metered pipeline sit fixed costs the vendor amortises into your rate: servers that hold the WebSocket open for the full call, orchestration that keeps latency under a second (a caller hangs up if the AI takes three seconds to respond), monitoring, and the engineering payroll behind all of it.

Add a margin and you get the retail price. When a platform charges $0.30–0.50 per minute, you are usually paying for one or more of: the most expensive TTS voices, a frontier-tier LLM on every turn, heavy sales-team overhead, or simply a generous margin because the buyer did not compare.

Why volume changes the economics — genuinely, not as a sales trick

Volume discounts in AI calling are not arbitrary. Three real cost curves bend as usage grows. Telephony: carriers sell minutes in tiers; a platform committing to millions of minutes buys each minute cheaper than one buying thousands. AI inference: cloud providers offer committed-use and batch pricing for ASR, LLM and TTS that can cut unit inference cost 30–50% versus on-demand rates — but only platforms with predictable volume can commit. Infrastructure utilisation: a server that handles 5 concurrent calls at 10% utilisation costs the same as one running at 80%; high-volume customers smooth the load and make every server-hour productive.

That is why honest volume pricing exists. As your monthly usage grows, the platform's blended unit cost genuinely drops — and a fair vendor passes it on. Our tiers reflect that curve directly: ₹6/min (about $0.07) up to 1,000 minutes, ₹4/min (about $0.05) from 1,000 to 2,000, then ₹3.8, ₹3.5 and ₹3.2 per minute through 5,000 minutes, and just ₹3/min (~$0.036) beyond 5,000 minutes a month.

Note what does not change across these tiers: the voices, the model quality, and the latency target are identical. The discount comes from the cost curves above, not from switching you to a worse stack.

The trap: cheap rates that cut quality where you cannot see it

There are two ways to hit a low per-minute price. The honest way is scale economics. The other way is degrading the pipeline: a robotic concatenative TTS voice instead of a neural one, a small quantised LLM that loses the thread after two turns, ASR without Indian-accent training, or routing through congested telephony that adds audible lag.

These cuts are invisible on a pricing page and very visible to your customers. Our internal testing is blunt about the thresholds: response latency above ~1.5 seconds makes callers talk over the AI; a voice that mispronounces Indian names erodes trust in the first ten seconds; and a model that cannot hold context across a three-minute conversation produces the "I already told you that" frustration that gets calls cut.

So when you evaluate any low rate — including ours — test the same three things on a real call: ask a follow-up question that depends on something you said a minute earlier (context), say a name like "Lakshminarayanan" and an address with a PIN code (ASR + TTS quality), and interrupt the AI mid-sentence (barge-in handling). Thirty seconds of testing reveals more than any pricing table.

A worked example: what 50,000 minutes a month actually costs

Take a mid-size operation — a real-estate developer's pre-sales team, a clinic chain's reminder programme, or a D2C brand's COD-confirmation flow — running 50,000 call minutes a month.

At a typical international platform's $0.30/min, that is $15,000 — roughly ₹12.75 lakh — every month. At a mid-market $0.10/min, about ₹4.25 lakh. On Nutaan's volume pricing, 50,000 minutes is far past the 5,000-minute mark, so every minute bills at just ₹3 — about ₹1.5 lakh, versus ₹8.5 lakh at $0.10 and ₹25 lakh+ at $0.30 for the same month.

For perspective against the human alternative: one lakh minutes is roughly the talk-time of 35–40 full-time telecallers (at a realistic ~45 hours of actual talk-time per agent per month). Their salaries alone exceed ₹35 lakh a month before training, attrition, infrastructure and management — and they work one shift, not twenty-four hours.

What we deliberately do not charge for

Pricing comparisons fail when the meter definitions differ, so here is ours precisely: we bill connected call time only, per minute. There is no platform fee to start (the free tier ships with ₹100 of credits), no per-agent licence, no setup fee, and no charge for failed or unanswered call attempts. Chat is metered separately (₹2 per AI message, or unlimited on the ₹2,400/month Pro plan) because chat's cost structure has no telephony component — bundling them into one opaque rate hides cross-subsidy.

The thing to watch on any platform, ours included: concurrency limits. A rate is meaningless if you are capped at five simultaneous calls when your campaign needs two hundred. Nutaan's campaign engine parallelises outbound batches; if you have specific concurrency needs at volume, that is an Enterprise conversation — and unlike the per-minute rate, it is negotiable.

How to run a fair evaluation

First, estimate your real monthly minutes: number of calls × realistic average duration. A COD confirmation averages 40–60 seconds; a lead qualification 2–4 minutes; a CCM check-in 3–5 minutes. Most teams overestimate duration and underestimate count.

Second, price three scenarios — your launch volume, 3× that, and 10× that — on each vendor's published tiers. The 10× number is where the pricing models diverge dramatically, and it is the number that matters if the channel works.

Third, run a same-script bake-off on real calls with the quality tests from earlier. Fourth, check exit friction: month-to-month billing and exportable data mean a vendor must stay cheap and good to keep you; annual lock-ins mean they only had to be convincing once.

You can run the entire evaluation on Nutaan without talking to anyone: sign up free, the welcome credits cover your first test calls, and the volume calculator on our pricing page shows your exact rate at any usage level. If your scale justifies it, the Enterprise tier adds white-label and custom volume terms — but the published ₹3/min floor at volume is already the rate most vendors save for a negotiated enterprise contract.

Written by the Nutaan team — we build and operate the AI calling and chat platform described here, running production voice and chat workloads for businesses across retail, healthcare, real estate, education and hospitality. Updated 12 June 2026. Pricing figures reference Nutaan's published rates; costs on other platforms vary. Questions or corrections: avishek@nutaan.com.