What an AI voice agent really costs per minute

Across the six platforms we work through below, a realistic all-in cost is roughly $0.05 to $0.31 a minute. The spread is that wide because most headline rates quote only the platform’s own fee; the parts it leaves out (the AI model, the transcription, the voice, the phone line) can take the real bill to six times the sticker.

That six-times figure is not a guess, and it is not the same for every platform. We derive it below, vendor by vendor, from the dated pricing data on our own vendor pages. Workings shown throughout.

The five parts of a minute

A voice-agent minute is really five bills stapled together. The platform fee pays the company running the call. Speech-to-text (STT) turns the caller’s words into text the software can read. The language model (the LLM, the AI brain) decides what to say back. Text-to-speech (TTS) reads that reply out in a voice. And telephony is the phone line itself. Some platforms bundle all five into one number; most quote one or two and meter the rest.

Retell itemises more cleanly than anyone, so its published rates make a good reference card:

(A SIP line just means you plug in your own phone-number supplier instead of renting the platform’s.)

The model menu moves the bill most

Look at that language-model row again, because it deserves its own table. This is the single choice that moves your bill more than the platform you pick.

Top to bottom, that menu spans a 53 times spread (0.16 ÷ 0.003) inside a single platform. Two things jump out. First, the floor is genuinely cheap: 0.055 infrastructure + 0.015 voice + 0.003 nano + a free SIP line adds up to about $0.073 a minute, which matches the bottom of Retell’s own banner. Second, do not assume the famous names sort the way you expect: Gemini 3.5 Flash at $0.081 costs nearly double GPT 4.1 at $0.045. A typical build (GPT 4.1 plus a US Twilio line) works out at 0.055 + 0.015 + 0.045 + 0.015, about $0.13 a minute before add-ons.

Four pricing shapes, worked through

Every platform we track prices in one of four shapes. Here is each one with real numbers.

One flat rate. Bland bundles the model, the listening, the speaking and the phone line into a single figure: $0.14 a minute on the free entry tier, $0.12 with a $299 monthly fee, $0.11 with a $499 fee. Predictable, which is the whole appeal. The catch is arithmetic: the $299 fee saves you $0.02 a minute, so it only pays for itself past 299 ÷ 0.02, roughly 15,000 minutes a month. Below that volume, the “dearer” $0.14 tier is the cheaper bill.

A platform fee plus pass-through. Vapi charges $0.05 a minute to host the call, and that is the only number Vapi sets; the other four parts are billed at cost by whichever suppliers you wire in. Honest, but only a floor: our stored band for a realistic stacked build runs $0.05 to $0.30. ElevenLabs works the same way for its agents at a dearer base: $0.08 a minute for the premium voice, your model on top, about $0.02 for a Twilio line, landing $0.08 to $0.25. Retell, worked above, is the semi-managed version of this shape.

A bundled all-in rate. Deepgram folds speech-to-text, the model and its own voice into one $0.075 rate; add a Twilio line at about $0.014 and you get 0.075 + 0.014, about $0.089 a minute (its premium tier is about 0.163 + 0.014, call it $0.18). Speechify goes further and includes the call routing too: from $0.07 a minute, with the pricing page promising “no passthrough, no token math”. That is the counter-model to everything above, though it is a young product we have not yet put through test calls.

Subscription plus overage. Speechify’s tiers also show the fourth shape: Pro is $99 a month with 1,200 agent minutes included, then $0.07 a minute after (overage just means the metered rate once your included minutes run out). Use all 1,200 and the effective rate is 99 ÷ 1,200, about $0.0825 a minute. Use only 600 and it is 99 ÷ 600, $0.165, double the rate for the same product. The allowance you do not use is the hidden cost of this shape.

One more line for the budget: the add-ons stack quietly. Vapi’s HIPAA mode (the healthcare data-handling standard) is a flat $2,000 a month on top of usage, and Retell’s automatic call-quality checks cost $0.10 a minute once you pass the first hundred free minutes. Neither shows up in a per-minute comparison until you read the footnotes.

The multiples, derived

You will see claims that voice agents “really cost 2 to 4 times the advertised rate”. We can do better than a folk figure, because we store both numbers for every platform: the vendor’s headline and a sourced all-in band. Divide the top of the band by the headline and you get each platform’s honest worst-case multiple.

(Deepgram’s rates were re-checked unchanged on 2026-07-11; the Retell all-in band is Retell’s own published banner, and the others are our sourced estimates of a realistic stacked build.)

Read the pattern rather than the individual numbers. The multiple is not a measure of vendor honesty; it is a measure of how much of the bill the headline includes. The pass-through platforms (6.0×, 4.4×, 3.1×) quote a floor and let your component choices set the rest. The fully bundled platforms (1.2×, 1.1×) quote nearly the whole bill, and at the bottom of their bands they come in under their own headline: Bland’s $0.11 top-volume rate sits below its $0.12 mid-tier figure, and Speechify’s $0.06 Enterprise floor undercuts its $0.07 sticker. Deepgram’s 2.4× lands in between because it bundles the engine but not the phone line, and its premium engine tier more than doubles the standard rate. So the fair summary is not “expect 2 to 4×”. It is: on a pass-through platform the worst case runs 3 to 6 times the headline and your choices decide where you land; on a fully bundled platform the headline is the bill, give or take about 15 per cent.

When each shape wins

Our position, since a table without a recommendation is just homework. If you want a number you can put in a budget and forget, buy bundled: Bland for phones at volume, Speechify if you want the cheapest bundled rate and can live with a young product and thin compliance paperwork. If you have a developer and a cost target, buy pass-through and shop the components: Retell’s $0.073 floor and Vapi’s $0.05 platform fee reward people who pick a budget model and bring their own line, with the trade-off that you now reconcile several suppliers’ bills and every saving is your own homework. If the voice itself is what sells your calls, pay ElevenLabs’ $0.08 base and accept the dearer top end; a premium voice on a bargain stack is usually the wrong way round. And if your volume is steady and predictable, a subscription with included minutes prices well, but only if you actually use the allowance you are paying for.

The honest limits

Four caveats, stated plainly. We have not yet run our own test calls, so nothing here says which platform sounds better or answers faster; this is a pricing piece, and price is one axis of several. Every figure is dated (captured between 15 June and 11 July 2026) and pricing genuinely moves: Speechify restructured its entire pricing between our June and July captures. Your prompt and model choice move the bill more than the platform choice, as the 53× menu spread shows, so treat every band here as a bracket rather than a promise. And these are estimates from published rates, not quotes; at enterprise volume most of these vendors will negotiate.

Next step: put your own monthly minutes into the cost calculator and see the six platforms above ranked for your usage, check any headline against its all-in band in the price index, or line up your two shortlist picks on the comparison pages. Each vendor page carries the full pricing notes with capture dates and screenshots.