Text-to-speech stopped being the expensive part of a voice agent this year. Google’s Gemini 2.5 Flash TTS bills $10 per million audio output tokens, which converts to roughly $0.015 a minute of finished audio — about 90 cents per audio hour, the cheapest credible conversational voice on the market. ElevenLabs, still the quality leader, cut its self-serve API prices on May 7: Flash from $0.11 to $0.05 per 1,000 characters (down 55%), speech-to-text from $0.40 to $0.22 (down 45%), and agent minutes from $0.10 to $0.08 (down 20%). Google’s pricing is on the Cloud Text-to-Speech pricing page; ElevenLabs documented its cuts in its own pricing announcement.
So the synthesis meter collapsed. And it changed almost nothing about what a voice deployment actually costs, because synthesis was never where the money went.
A cheap call is not a cheap system
Operators now put the all-in cost of an AI-handled contact between $0.20 and $2.00, with a median near $0.90, against roughly $13.50 for a human-assisted contact. Run the arithmetic on a four-minute call. At the Gemini floor, four minutes of synthesized audio is about $0.06. Even on ElevenLabs’ premium multilingual tier — $0.10 per 1,000 characters, which the vendor’s own API pricing page equates to roughly $0.10 a minute — the same call carries about $0.40 of voice.
That means voice generation is somewhere between 7% and 44% of a task that costs ninety cents on median. When the top-of-market model is roughly five times the floor per audio hour and the cheap model already sounds acceptable, the remaining 56% to 93% of the invoice is where every negotiation should start.
Three meters, one call, a 24x spread
The reason voice budgets stay murky is that every vendor publishes a different unit, and none of the published units describe the same thing. Plura AI’s September unit-economics breakdown normalizes published pricing into cost per contact at 100,000 monthly contacts and a four-minute average handle time:
- Amazon Connect at $0.038 per voice minute becomes $0.152 per contact — voice service only, with telephony, Lex and conversation analytics billed separately.
- Five9 at $159 per concurrent seat per month with a 50-seat minimum lands at $0.0795 per contact for the license alone.
- Genesys at $75 per named user per month lands at $1.875 per contact, also license only.
Same four-minute call, three published units, a 24x spread — and the AI itself is not in any of those numbers. This is the practical consequence of the pricing-model split we covered in Voice AI Pricing: From Seats to Minutes to Outcomes: per-seat, per-minute, per-resolution and per-outcome rates are not four versions of one price, they are four different products with four different risk allocations. Convert every quote to cost per completed conversation before you compare anything, and write the conversion assumptions — handle time, containment rate, concurrency — into the proposal so the client can audit them later.
The line items nobody puts on the rate card
Four costs dominate a real deployment and appear in almost no published price:
- Telephony. Phone numbers, carrier minutes, and the SIP or WebRTC topology under the agent. The carrier layer is measurable and often the difference between a demo and a usable agent.
- Peak concurrency. Voice traffic is bursty; Monday morning volume can run several times the weekly average. You provision for the peak and pay for it all month.
- The realtime turn. Speech-to-text, the model, and speech-to-speech generation each have to fit inside a conversational latency budget. Transcription is no longer the weak link: the best streaming models now sit near 2.5% word error rate, and the accuracy-latency frontier runs at 26 to 45 milliseconds median finalization on independent rolling benchmarks.
- Evaluation and human review. Listening to calls, scoring transcripts, and maintaining a regression set is labor, and it is the only thing that keeps containment from drifting.
Those plumbing costs are also the most common reason a pilot never scales, which is why a pre-launch readiness audit of the network and hand-off points belongs in the budget before the prompt work starts — the checklist we laid out in Voice AI Readiness: The Pre-Launch Gate Most Pilots Skip.
The leak that dwarfs every model swap
A failed handoff bills twice. When the agent escalates mid-call, the operator pays for the AI minutes already spent plus the human contact that finishes the job. Containment rate, not cost per token, is the number that moves the budget.
Take a deliberately conservative illustration. Assume 100,000 monthly contacts, a $0.90 AI cost per contact, and failed handoffs that add a three-minute human contact at $0.50 a minute. Each failure wastes about $2.40 — the $0.90 already spent plus $1.50 of human time. Lifting containment from 60% to 75% removes 15,000 of those failures: roughly $36,000 a month. No voice model price cut, at any tier, comes close to that.
What to do with the savings
Three moves follow from the arithmetic.
Re-price the offer on the resolution, not the minute. If synthesis is a rounding error, per-minute pricing exposes you to a cost you no longer control the margin on and hides the value you actually deliver. Anchor the proposal to completed conversations or resolved intents, then state the containment assumption in writing.
Spend the synthesis savings on the operations layer. Co-locate speech-to-text, the model, and speech generation in the same region as the SIP gateway, benchmark carriers, and stand up a transcript review loop. That spend shows up as containment; a voice upgrade mostly shows up as a nicer demo.
Match voice tier to revenue. Keep premium expressive voices where prosody carries money — brand moments, high-value outbound, character work. Use the commodity tier for menus, confirmations, and status calls. This is the same containment-versus-cost tradeoff that decides whether a business keeps its IVR at all, as we argued in Voice AI vs IVR: Why Call Centers Are Switching in 2026.
The floor for synthesized speech will keep falling. The invoice a client actually pays will keep being decided by telephony, concurrency, and how often the agent finishes the job without handing it to a human. If you are designing voice work for clients and want the offer priced against the economics rather than the meter, our team does that as part of digital marketing and AI service design.