What an AI Phone Agent Costs Per Minute, Broken Down

TL;DR

• Advertised platform fees ($0.05–$0.07/min) cover only the orchestration layer — not the speech, the model, or the phone line.
• A mid-tier production stack typically lands between $0.12 and $0.25 per minute all-in, rising past $0.30 with premium voices or frontier-class LLMs.
• TTS is usually the biggest single line item, not the LLM — voice choice matters more than model choice for most conversational agents.
• BYOK (bring-your-own-key) platforms give you component-level visibility; bundled platforms trade that visibility for a simpler invoice, usually at a 15–40% markup.
• Hidden costs — silence billing, concurrency overage, HIPAA surcharges — can add 15–30% on top of the per-minute estimate you build in a spreadsheet.

---

The number vendors advertise is not the number you pay. The advertised rate for AI voice agent platforms ranges from $0.05 to $0.15 per minute, but the actual all-in cost is $0.12–0.25 per minute for most production deployments — the gap exists because advertised rates typically reflect only the platform fee, not the STT, LLM, TTS, and telephony costs layered on top. That gap is not a trick; it reflects a genuinely modular architecture. Understanding each module is how you build a budget that survives contact with a real bill.

What is actually happening during one minute of a call?

Every AI phone call chains four services in real time. Most AI voice agents charge for multiple layers: the platform or orchestration fee, speech-to-text (STT) to convert caller audio to text, a large language model (LLM) like GPT-4.1 or Claude to generate responses, text-to-speech (TTS) to convert those responses back to audio, and telephony to connect the actual phone call. The platform you pick determines whether you see those costs itemised separately or collapsed into one per-minute number.

The orchestration layer is the smallest bill. The $0.05/min platform fee on something like Vapi pays for the real-time pipeline, endpointing, barge-in, call routing, dashboards, transcripts, recordings, and observability — it does not include STT, LLM, TTS, or the phone-number carrier. That honesty is real, but it also means the advertised number tells you almost nothing about your actual cost.

Layer 1 — Speech-to-Text (STT)

STT is the cheapest layer. When the caller speaks, the audio signal must be converted to text in milliseconds for the AI to understand it — leading models like Deepgram, Whisper, or Azure charge approximately $0.01–$0.02 per minute for this.

Deepgram is the most common choice for real-time voice pipelines. Deepgram's cheapest speech-to-text rate is $0.0043 per minute for pre-recorded Nova-3, while streaming starts at $0.0048 per minute, a limited-time promotional rate against a regular price of $0.0077 per minute. For a voice agent, you need streaming, not batch — the latency difference is the whole product. Batch processes a recorded file in minutes; streaming gives you words as they are spoken. At the promotional streaming rate, STT rounds to under half a cent per minute, which makes it easy to under-weight in a budget. The thing to watch is add-ons: enabling diarization on streaming audio, for example, can more than double the effective per-minute STT cost.

Layer 2 — The LLM

The LLM is where cost variance is highest, because model choice is entirely up to you. The LLM is the "brain" of the agent: it processes transcribed text, understands context, and generates the appropriate response — models like OpenAI's GPT-4o, Anthropic Claude, or Google Gemini charge by tokens.

LLMs are priced per token, and per-minute estimates should apply a reality multiplier to capture conversation-history token growth (which compounds with turns), function-calling round-trips, and barge-in handling. A system prompt that runs 2,000 tokens gets prepended to every turn. That compounds fast on a long call. An average one-minute phone call might involve somewhere between 500 and 2,000 tokens depending on conversational complexity, the length of the system prompt, and how verbose the caller is — at those volumes the model layer costs roughly $0.001 per minute on a smaller model and around $0.02 per minute on a larger one, a twentyfold spread driven entirely by model choice.

The practical implication: for a simple appointment-booking agent with a tight script, a smaller and cheaper LLM is defensible. For an agent that needs to reason across a knowledge base, handle objections, or integrate with backend systems mid-call, reaching for a frontier model is often correct even at five to ten times the per-token cost. Match the model to the task before you model the cost.

Layer 3 — Text-to-Speech (TTS)

TTS is almost always the largest single line item in the stack, and it is the one most cost estimates under-represent. Text-to-speech can introduce unexpected cost escalations, especially when teams opt for premium or neural voices that deliver more natural human-like conversations — these upgraded voices often carry higher per-character or per-minute rates.

ElevenLabs is the common premium choice. ElevenLabs' API pricing page lists $0.10 per 1,000 characters for Multilingual v2 and v3, and $0.05 per 1,000 characters for Flash and Turbo. At about 900 characters per spoken minute (150 words), Multilingual v2 costs about $0.09 per minute and Flash about $0.045 per minute at those API rates. That alone is already larger than the platform orchestration fee. The gap between a Flash voice at $0.045/min and a Multilingual v3 voice at $0.09/min is real money at volume — and most teams pick the better-sounding voice without running the math first.

For latency-sensitive pipelines, Deepgram's own TTS is worth evaluating. Deepgram Aura-2 provides ultra-low latency at $30 per million characters — and if latency is the priority, Deepgram's stack has an advantage. That works out to roughly $0.027 per minute of speech, cheaper than ElevenLabs Flash and meaningfully faster on time-to-first-audio, which affects how natural the conversation feels.

Layer 4 — Telephony

Telephony is the most predictable layer. Telephony is the fundamental infrastructure for making or receiving calls over the telephone network — providers like Twilio, Telnyx, or local SIP trunk providers charge fees for numbers and connection minutes, typically $0.01–$0.03 per minute. The number to watch is not the per-minute rate; it is the regulatory pass-through. Telephony figures typically exclude US 5–15% USF and regulatory pass-through fees, plus the monthly per-DID rental cost. For a US deployment with ten phone numbers, DID rental alone adds a fixed overhead before a single call is placed.

How does the platform model change the math?

The choice between BYOK and bundled platforms does not change the underlying component costs — it changes how visible they are. BYOK platforms charge a low orchestration fee ($0.04–$0.07/min) and let you connect your own STT, LLM, TTS, and telephony accounts, giving maximum control but requiring you to manage multiple vendor relationships; bundled platforms package everything into one per-minute rate ($0.11–$0.50/min), offering simplicity at the cost of flexibility and typically a 15–40% markup over component costs.

The right answer depends on your team's engineering capacity and call volume. The crossover point is usually around 20,000 minutes per month — below that, a managed or bundled setup pays for itself in engineering time saved. Above that threshold, the markup compounds enough that BYOK starts making economic sense, provided someone owns the operational overhead of managing four separate vendor relationships.

Cost Layer	Representative Provider	Low ($/min)	High ($/min)	Biggest lever
Platform / Orchestration	Vapi, Retell, Bland	$0.05	$0.14	BYOK vs bundled
STT	Deepgram Nova-3 streaming	$0.005	$0.02	Add-ons (diarization, redaction)
LLM	GPT-4o mini to GPT-4o	$0.001	$0.08+	Model tier and prompt length
TTS	Deepgram Aura-2 to ElevenLabs v3	$0.027	$0.09	Voice quality tier
Telephony	Twilio, Telnyx	$0.01	$0.03	Regulatory pass-through, DID count
All-in (mid-tier stack)		$0.12	$0.30+	Voice and model choice dominate

Rates above are drawn from publicly available 2026 pricing and sources cited in this article. Verify each provider's current rate card before building a budget — 2026 has already seen multiple rounds of LLM price cuts, and provider rates move regularly.

What costs hide behind the per-minute number?

Hidden fees like silence billing, concurrency limits, and HIPAA surcharges can add 15–30% to your per-minute estimate. Silence billing — charging for seconds when neither party is speaking — is a specific gotcha on some platforms. Concurrency limits matter if your call volume spikes: on Vapi's Build plan, concurrency past the included lines costs $10 per concurrent line per month, and a HIPAA compliance add-on is substantial — for a regulated team running 5,000 minutes a month, the compliance overhead alone can cost more than the calls.

The other underestimated cost is the engineering work to keep the agent reliable after deployment. Prompt changes, model version bumps, and new edge cases all need testing. That is not a per-minute charge, but it is a real cost of operating a voice agent in production — and it does not appear on any vendor invoice.

Which pricing model fits which situation?

Per-minute billing and per-resolution billing serve different shapes of business. Per-minute pricing runs $0.05–$0.30 per minute for the AI plus $0.005–$0.05 per minute for telephony — it is better for short calls and worse for variable-length conversations. Per-resolution pricing charges a flat fee only when the agent fully resolves the caller's issue, typically $0.70–$1.50 per resolved call depending on vendor. That model transfers duration risk to the vendor, which is attractive for customer-service workloads where call length varies widely. The trade-off is that the definition of "resolved" matters enormously — ask vendors to show you exactly how resolution is counted before signing anything.

If you are evaluating platform options side-by-side, the comparison article Vapi vs Retell vs Bland: Choosing a Voice AI Platform in 2026 walks through the architectural differences. For context on how LLM token costs behave at scale across any kind of agent, LLM Cost Optimization: Caching, Routing, and the Token Waste Hiding in Your Prompts covers the levers that apply whether you are building a voice agent or a text-based one.

For the engineering work of scoping, building, and operating a voice agent — including deciding which stack actually fits your volume and compliance requirements — see my voice AI agents service page. The Biz365 AI case study shows the kind of integrated AI pipeline work that underlies this type of build. Most engagements start around $5K; smaller well-scoped work is considered case-by-case.

---

FAQ

What is a realistic all-in cost for an AI phone agent per minute in 2026?

Production deployments typically cost $0.12–0.25 per minute once you stack speech-to-text, LLM inference, text-to-speech, telephony, and platform fees — with the five cost layers being STT ($0.004–0.024/min), LLM ($0.003–0.08/min), TTS ($0.02–0.10/min), telephony ($0.008–0.014/min), and the platform orchestration fee ($0.05–0.14/min). Premium voice models and frontier LLMs can push the total past $0.30/min.

Is TTS really more expensive than the LLM for a typical voice agent?

For most conversational agents — appointment booking, intake forms, FAQ handling — yes. A premium ElevenLabs Multilingual v3 voice costs roughly $0.09/min in TTS alone. A mid-tier LLM on a focused task might cost $0.005–0.02/min in tokens. The premium voice costs four to eighteen times more than the LLM layer. TTS cost is driven by voice quality tier, not call volume tier, so it surprises teams who optimise the wrong thing first.

Does BYOK always save money compared to a bundled platform?

The crossover point is usually around 20,000 minutes per month — below that, managed or bundled platforms tend to pay for themselves in engineering time saved. Above that threshold the 15–40% bundled markup becomes material. The hidden cost of BYOK is managing four separate vendor relationships, four separate rate cards, and four separate failure modes. That is real engineering overhead, especially when one provider has an incident mid-campaign.

How does HIPAA compliance affect cost?

Compliance adds a fixed overhead that is independent of call volume. Compliance is not included in Vapi's base pricing. For any healthcare or regulated deployment, get the compliance cost itemised separately from the per-minute rate before comparing platforms — a platform that looks cheaper per minute can be more expensive once compliance infrastructure is factored in.

What is the single biggest cost lever available to me?

Voice choice. The gap between a Flash-tier TTS voice and a premium multilingual neural voice can be two to four cents per minute. At 10,000 minutes per month, that is $200–$400 per month from a single configuration decision. Evaluate whether a lower-tier voice is perceptibly worse to your specific callers before defaulting to the most realistic voice available. In many B2B contexts, callers care more about response accuracy and latency than voice naturalness.