Every vendor sells you a number per minute. Almost none of them tell you what's actually inside it. Here's the full unbundled cost — TTS, STT, LLM, telephony, infra — for an AI voice agent in 2026.
Every voice-AI vendor pitches you a single number: $0.12/min. $0.30/min. $0.99/min. The number sounds clean, but it's a black box. Inside that minute is a stack of four to six paid services, each metered differently, each with margin baked in. If you don't know what's actually being billed, you can't tell whether you're getting a fair deal — or whether you should just build it yourself.
Here's the unbundled 2026 cost of one minute of an AI voice agent on a phone call. Real numbers, real providers, current as of Q2 2026.
A voice agent on a phone call is five services running in parallel, glued together by orchestration code. Strip away the marketing and every vendor is reselling some combination of these:
There's a sixth line item nobody talks about: failed calls and silence. You pay for telephony minutes whether or not the agent says anything useful.
Assume an average call: 45 seconds of caller speech, 30 seconds of agent speech, ~600 LLM tokens in / ~300 out per turn, 4 turns per call, 60-second total billable minute.
| Component | Provider | Rate (May 2026) | Cost / min |
|---|---|---|---|
| Telephony (inbound US local) | Twilio | $0.0085/min | $0.009 |
| STT (streaming) | Deepgram Nova-3 | $0.0043/min | $0.004 |
| LLM | GPT-4.1-mini, ~2,400 in / 1,200 out per call | $0.40/$1.60 per M tok | $0.003 |
| TTS | ElevenLabs Flash v2.5, ~600 chars | $0.10 per 1k chars | $0.060 |
| Orchestration | Self-hosted Pipecat on a $40/mo VPS, ~5k min capacity | infra-amortized | $0.008 |
| Total | ~$0.084 / min |
That's the floor. About 8.4 cents a minute if you assemble it yourself with mid-tier components.
Now compare to what hosted platforms charge:
The gap between DIY ($0.08) and a polished managed platform ($0.15–0.25) is your convenience tax. For under 50,000 minutes/month, the convenience tax is probably worth it. Above that, you're leaving real money on the table.
People obsess over the wrong line. They argue about which LLM to use. The LLM is the cheapest part of the stack. At ~$0.003/min it's literally a rounding error.
The two costs that actually matter:
TTS dominates. ElevenLabs Turbo/Flash is ~70% of the variable cost in our breakdown. You can cut this in half by switching to Cartesia Sonic ($0.04/min equivalent) or Rime, but you trade voice quality. For a brand-sensitive use case (luxury hotels, high-end legal) the ElevenLabs premium is worth it. For HVAC after-hours triage? Cartesia is fine.
Telephony adds up at scale. $0.009/min sounds tiny until you're doing a million minutes/month. That's $9k just to keep the lines open. Negotiate volume rates with Twilio at 250k+ minutes — typically 20–35% off rack rate. Telnyx is ~25% cheaper than Twilio out of the box but has more carrier-routing edge cases.
Three line items that don't show up in any vendor's pricing table but always show up in your bill:
Add ~12–15% on top of your "happy path" cost to model these realistically.
The question isn't "what does an AI voice agent cost in 2026." The question is: what's the gross margin between what you pay per minute and what your customer pays you per appointment, lead, or recovered no-show?
A 60-second call that costs you 8 cents and books a $4,200 HVAC repair has 99.998% gross margin. That math doesn't change whether you're paying $0.08/min or $0.25/min. So pick the stack you can actually ship and ship it. Optimize the unit economics later — when you have units.