The Real Cost of a Finished Minute of AI Voice
$/finished minute, not $/1M chars the number that matters here
By Stan K.
The same finished hour of narration costs $0.81 on OpenAI's tts-1 and $12.40 on an ElevenLabs Starter plan. We converted every verified 2026 TTS price — characters, credits and audio tokens — into dollars per finished minute, regeneration tax included.
TTS pricing pages are written in three currencies nobody listens in: characters, credits and audio tokens. ElevenLabs bills characters (1 credit = 1 character), Google bills audio tokens (25 per second), OpenAI offers both depending on the model. Your listener hears minutes. Every number below is a verified 2026 price converted into the only unit that matters — what one finished, usable minute costs — and then what a full hour of shipped audio actually runs, because a finished minute is never the last minute you generate.
The conversion basis, so you can check our math
| Vendor unit | Official rate | One finished minute equals |
|---|---|---|
| ElevenLabs credits | 1 credit = 1 character | ~750 characters |
| Cartesia credits | ~750–800 credits per minute of audio | ~775 credits |
| Google audio tokens | 25 tokens per second | 1,500 tokens |
The ~750 characters per minute anchor comes from Cartesia’s official meter — the one vendor that publishes a credits-to-minutes rate. One row refuses conversion: OpenAI’s gpt-4o-mini-tts bills $12 per 1M audio output tokens but publishes no token-to-audio-second rate, so its per-minute cost is unverifiable. We leave it as ”—” rather than guess.
Every verified price, per finished minute
Sorted cheapest first — the order surprises people:
| Provider / tier | Sticker price | $ per finished minute |
|---|---|---|
| Kokoro-82M (local, Apache 2.0) | $0 | $0.00 |
| OpenAI tts-1 (API) | $15 / 1M chars | $0.011 |
| Gemini 3.8 Flash TTS (API, promo) | $9 / 1M output tokens | $0.014 — doubles to $0.027 after 2026-12-31 |
| Gemini 2.5 Flash Preview TTS (API) | $10 / 1M output tokens | $0.015 |
| OpenAI tts-1-hd (API) | $30 / 1M chars | $0.023 |
| ElevenLabs API, v4 Turbo | $40 / 1M chars | $0.030 |
| ElevenLabs API, v4 | $80 / 1M chars | $0.060 |
| Cartesia Pro (in quota) | $5 / 100k credits | ≈$0.04 (overage ≈$0.05) |
| ElevenLabs Creator (in quota) | $22 / 121k credits | ≈$0.14 (overage $0.18) |
| ElevenLabs Starter (in quota) | $6 / 30k credits | ≈$0.15 (overage $0.20) |
| OpenAI gpt-4o-mini-tts (API) | $12 / 1M audio tokens | — (no published rate) |
The row that should stop you: ElevenLabs Starter credits cost ~$0.15 per minute while the same vendor’s v4 Turbo API costs $0.030. The subscription charges five times the API price for the same audio — you are paying for the studio, the convenience and the included quota, not for minutes. And ElevenLabs’ free tier (10k credits ≈ 13 minutes) forbids commercial use outright in the Terms, so a commercial hour legally starts at Starter — that trap and others are dissected in Seven TTS Fine-Print Traps.
The regeneration tax: 60 shipped minutes are never 60 generated
Nobody ships the first take of everything. A mispronounced name, a wrong emphasis, a sentence rewritten after the voiceover existed — every scrapped take was still billed. If one segment in five gets regenerated, your generated total is 1.2× your shipped total.
60 shipped min → 72 generated minthe workload behind every number below
One finished hour at a 1-in-5 regeneration rate bills 54,000 characters (72 min × 750), 108,000 Google audio tokens (72 × 1,500) or ~55,800 Cartesia credits (72 × ~775). Change the regeneration rate and every total scales linearly — the ratios between providers do not move.
What one finished hour actually costs
| Provider / tier | The bill for that hour |
|---|---|
| Kokoro-82M, local | $0.00 — the electricity a CPU box draws anyway |
| OpenAI tts-1 | $0.81 |
| Gemini 3.8 Flash TTS (promo rate) | $0.97 — $1.94 after 2026-12-31 |
| Gemini 2.5 Flash Preview TTS | $1.08 |
| OpenAI tts-1-hd | $1.62 |
| ElevenLabs API v4 Turbo | $2.16 |
| ElevenLabs API v4 | $4.32 |
| Cartesia Pro | $5 flat — ~56k of 100k credits used, room to spare |
| ElevenLabs Starter | $12.40 — $6 plan + ~$6.40 overage (quota covers only ~40 min) |
| ElevenLabs Creator | $22 flat — 54k of 121k credits (a single-hour month wastes the rest) |
Three reading notes. Flat rows assume this hour is your only usage that month — spread over a busy month, Creator and Cartesia Pro amortize much better than the single-hour column suggests. The Gemini row is a promo rate with an expiry date stamped on it; budget the post-2026-12-31 number. And the spread itself is the story: $0.81 to $12.40 for the same hour of audio, a 15× difference between two reputable vendors — decided entirely by which unit you were quoted in.
The $0 row is not a rounding error
Kokoro-82M is an Apache-licensed, 82M-parameter model that voices our production English narration without touching the GPU — measured at roughly 20× faster than real time on CPU. At one finished hour per week, the cheapest paid row on this page still adds up to real money over a year; Kokoro adds up to watts. How we run it, and where local honestly stops, is in What Years of Local TTS Taught Us and One GPU Beats Ten Subscriptions.
All tiers, Elo quality references and license flags for every provider in this table live in the TTS hub.