Cost
What a voice agent actually costs per minute
Five meters, real published rates, and a volume threshold below which you should just buy a platform. The saving at scale is real, but it is smaller than people assume.
5 min read
Almost every voice AI vendor quotes a single number: so many rupees, or so many cents, per minute. It is a useful number and it is never the number you pay.
Below is the full model, built from published rates and worked across five carriers and two markets. It arrives somewhere more interesting than a figure: a volume threshold, below which assembling your own agent costs more than buying a bundled platform, and above which it saves you rather less than the pitch decks suggest.
The five meters
A voice agent is a loop, and each stage bills separately.
Telephony carries the call, charged per connected minute, with a monthly rental for the number and — in India — often a charge per concurrent channel.
Compute runs the agent itself. Charged per minute the agent is live, and also per minute an instance sits warm waiting for a call. This is the meter almost nobody models, and it is frequently the largest.
Speech-to-text bills per minute of audio and runs for the whole call, not only the parts where somebody is talking. Modelling it against the agent's speaking time is a common error.
The language model bills per token in and out. This is the line everyone expects to dominate.
Text-to-speech bills per character of what the agent says.
The rates
Published pricing pages except the Exotel figures, which are quoted directly because Exotel publishes no inbound rates. Checked September 2026.
| Component | Vendor and tier | Rate |
|---|---|---|
| Compute | Pipecat Cloud agent-1x, active | $0.010 / min |
| Compute | Pipecat Cloud agent-1x, warm and idle | $0.0005 / min |
| Speech-to-text | Deepgram Nova-3 streaming, multilingual | $0.0058 / min |
| Text-to-speech | Deepgram Aura-2 | $0.030 / 1k characters |
| Language model | Claude Haiku 4.5 | $1.00 / MTok in, $5.00 / MTok out |
| Telephony | Exotel, India | ₹0.20 / min + ₹499 number + ₹1,200 per channel |
| Telephony | Plivo, India | ₹0.38 / min + ₹200 number |
| Telephony | Plivo, US | $0.0055 / min + $0.50 number |
| Telephony | Twilio, US | $0.0085 / min + $1.15 number |
| Telephony | Twilio, India | $0.0699 / min + $1.15 number |
Deepgram's multilingual rate is promotional; standard is $0.0092, so budget that. Conversions use ₹88 to the dollar.
The assumptions
Per call — three minutes, of which the agent speaks about 1.2, producing roughly 1,100 characters across 6 turns. The model sees a 1,200-token system prompt plus growing history, averaging about 1,900 input tokens per turn with 60-token replies. Derived from a conversation of this shape, not measured from one deployment.
Per month — 600 calls, 1,800 connected minutes, and one warm compute instance. That last one deserves flagging: a warm instance accrues about 1,440 idle minutes a day whether or not the phone rings, so the number you keep ready moves the total more than any other single setting. One is the assumption throughout. Two roughly doubles the compute line.
What one call costs
Cost of one three-minute call · ₹
- Telephony
- Compute (incl. idle)
- Text-to-speech
- Speech-to-text
- Language model
Compute at ₹5.68 is the largest line for every carrier except Twilio India — and more than half of it is idle time. You are paying mostly to keep the agent ready, not to hold the conversation. Dropping to zero warm instances saves ₹3.04 a call and costs you cold-start latency on the first response, which a sub-second budget cannot absorb. That is a real and quantified trade between cost and answer speed.
Should you build this at all?
Now the question that matters. A bundled platform in India retails around ₹6 per connected minute, all in. Here is the same stack, on the cheapest sensible carrier, at increasing volume.
Build your own vs a bundled platform · ₹ per minute
Vertical line: a bundled platform at ₹6 per connected minute. Bars to the right of it cost more than buying.
Below roughly 720 minutes a month — about eight calls a day — buying is cheaper. Warm compute and number rental are spread across too few minutes, and no amount of careful vendor selection fixes that. Above it, building wins, settling at about ₹3.72 a minute, roughly 38% under platform pricing, and it stops improving beyond a few thousand minutes because almost everything left is variable cost.
Thirty-eight percent is real money — at 18,000 minutes a month it is around ₹41,000 — but it is not the order-of-magnitude difference people expect when they set out to avoid a vendor. If your entire case for building rests on cost, the case is thinner than it looks.
Three more things the numbers say
Carrier choice is the biggest variable lever, but not the biggest number. The spread across carriers is 2.3×, from ₹4.25 on Plivo India to ₹9.97 on an international carrier terminating Indian calls. Worth getting right. Just not worth getting right before you have answered the volume question.
A lower per-minute rate can cost you more. Exotel charges ₹0.20 a minute against Plivo's ₹0.38 — less than half — yet at 600 calls a month works out 1.56× more expensive, because ₹1,200 per channel is owed whether the channel is busy or idle. The two cross at around 5,000 calls a month. Anyone comparing Indian carriers on the headline rate alone will choose wrong at one end of that range.
India is not the expensive market. On like-for-like carriers it is marginally cheaper than the US. The ₹9.97 figure is not India being costly; it is an international carrier terminating calls where it has no local infrastructure. Those are different facts and they get conflated constantly.
Building your own version
- Your carrier's inbound rate, and whether they bill per second or round up to the minute.
- Every fixed charge — number rental, channel fees, warm instances, platform minimums — divided by your actual monthly volume, not your hoped-for volume.
- Average connected call length, from call detail records.
- Characters of speech per call. If it is not instrumented, take average words and multiply by about six.
- Input and output tokens per call, summed across turns. A 4,000-token system prompt over eight turns is 32,000 input tokens on one three-minute call.
- How many instances you keep warm, and what cold-start latency costs you if the answer is none.
What it is for
The honest conclusion is not one a vendor of either kind will tell you. Below a few hundred calls a month, buy the platform. Above it, building saves something like a third — worthwhile, but not the reason to do it.
The reasons to build are that you own the conversation design, you keep your call data, and you can change the parts that do not fit your business. Those are good reasons. The cost saving is a bonus that arrives only after you clear the threshold, and anyone selling you a build on economics alone has not run this arithmetic.
We work out whether a voice agent fits your business, build it if it does, and hand it over so you own it. If you are weighing this up, a discovery call is thirty minutes and costs nothing.
Book a discovery call