Media-Minute Pricing: Examples & Companies

What is it

Media-Minute Pricing is a billing unit where customers are charged per minute of audio or video processed — used by speech, voice, and video AI vendors.

A media minute is the duration unit of AI that hears, speaks, or generates moving pictures. Where text models meter tokens and infrastructure meters GPU-hours, speech and video products meter the length of the media — a minute of audio transcribed, a minute of speech synthesized, a minute of a WebRTC call carried, a minute of conversational video rendered. The reason is structural: audio and video have no natural token boundary the buyer can count in advance, but they always have a runtime. A 12-minute support call, a 30-second ad, a one-hour podcast — each carries an obvious, estimable duration that maps closely to the compute required to process it.

The unit is shared across products that look very different on the surface. Deepgram, Speechmatics, and Gladia bill speech-to-text by the minute or hour of input audio. Retell AI, Vapi, and Synthflow bill voice agents by the connected minute of a phone call. Real-time media platforms like Daily and LiveKit bill per participant-minute of a call carried. On the video side, Twelve Labs bills video understanding by the minute of source footage, while Tavus and Luma AI bill generated video by the minute or second of output. The same root word — “minute” — covers transcription, synthesis, agents, transport, and generation.

What makes the unit interesting is the spread. The cheapest media minute in this corpus, raw WebRTC media on LiveKit, costs about $0.0005; the most expensive, a full unbundled voice-agent minute on Retell AI, reaches $0.31 — over 600x higher for the same sixty seconds. Between them sit machine transcription (Gladia at $0.144/hr, roughly $0.0024/min), video indexing (Twelve Labs at $0.042/min), and healthcare-agent time (Hippocratic AI at ~$9/hour, or $0.15/min). The minute is one unit; the price is a function of what happens during that minute. See choosing the right usage metric for why duration is the natural fit here.

How it works

The core formula is simple: media cost equals the per-minute rate for the chosen model, multiplied by the minutes of audio or video processed (usually metered to the second and rounded up, often with a short minimum). The complexity lives in the dimensions wrapped around that minute — which task, which model, real-time versus batch, accuracy tier, whether the meter is per participant or per stream, and whether the vendor exposes the minute directly or hides it behind credits.

The display unit is frequently a presentation choice rather than the meter. Gladia and Sarvam AI quote per hour of audio ($0.144/hr and ₹30/hr respectively) while Twelve Labs re-expresses its $0.042/min index rate as $2.50/hour — the same rate, two labels. Higher up the stack, Synthesia, Pika, Luma AI, and Opus Clip sell a credit pool that converts to minutes — the minute is the real value metric, but the buyer transacts in credits. Resemble AI goes finer still, billing per second of audio ($0.0005/sec TTS) rather than per minute.

Unit math: Carrying a 4-person, 30-minute video call on Daily is 4 × 30 = 120 participant-minutes; after the 10,000 free monthly minutes that is 120 × $0.004 = $0.48. Indexing a 60-minute video on Twelve Labs’ Marengo ($0.042/min) costs 60 × $0.042 = $2.52. A 1,000-minute/month outbound voice-agent campaign on Retell AI at $0.115/min composed cost is 1,000 × $0.115 = $115. Staffing an AI nurse on Hippocratic AI for 40 active hours ($9/hr) is roughly $360 — versus ~$39/hour for a human RN.

Because the meter tracks duration, the same lever — commitment and volume — discounts it across vendors. Daily auto-discounts the participant-minute from $0.004 down to $0.0015 (about 63% off) at 50M+ minutes/month; Gladia’s Growth tier prepays for commitment-based volume discounts; Resemble AI offers enterprise volume discounts up to 80% off its per-second rate. This per-minute discounting is the substance of the voice-API minute-billing trend — see also the introduction to usage-based pricing for the broader frame, and the ElevenLabs pricing calculator to model a media-minute bill directly.

Companies using this

Forty companies in the corpus meter media minutes. They cluster into four groups: transcription and speech-to-text APIs (Deepgram, Speechmatics, Rev AI, Gladia, Sarvam AI, Otter.ai, Fireflies.ai); voice agents and real-time infrastructure (Retell AI, Vapi, Bland AI, Synthflow, PolyAI, Parloa, Daily, LiveKit, Hume AI, Hippocratic AI); text-to-speech and dubbing (ElevenLabs, Murf AI, Resemble AI, WellSaid Labs); and AI video generation and understanding (Synthesia, Tavus, Hedra, Twelve Labs, Luma AI, Pika, Opus Clip, VEED AI, InVideo AI, Kaiber, Wonder Dynamics, Creatify, Descript, Krisp, Kustomer, Fal).

Patterns observed

• The minute is one unit, but the price encodes the work inside it. Twelve Labs makes it explicit within one product: Marengo indexing is $0.042/min while Pegasus analysis is $0.0292/min on the same footage. Hume AI ladders empathic voice from about $0.02/min to $0.072/min depending on model. Across the corpus the same sixty seconds spans over 600x (see the definition above); the duration is constant, and the per-minute rate is where product differentiation lives.

• Voice agents have converged on the unbundled per-minute stack. Retell AI sets the template (infra + LLM + TTS + telephony, composing to $0.07–$0.31/min with “no platform fee” — the component rates are in the table above). Vapi mirrors it with $0.05/min hosting plus at-cost pass-through, and Synthflow advertises the same pay-only-for-the-composed-minute model. The transparency is itself a lever: the buyer can audit exactly which per-minute components they are paying for.

• The metering base is not always the minute the buyer pictures. Daily meters per participant-minute, so a 4-way call bills four times a 1-way call of the same length, while Resemble AI bills per second and other vendors quote per hour (the display-unit games are detailed under “How it works”). Normalize to a single base — a stream-minute, a participant-minute, a second — before any cross-vendor comparison holds.

• Video vendors hide the minute behind credits more often than audio vendors. Synthesia sells credits that convert to video minutes, Opus Clip sets roughly one credit per minute of processed video, and Pika, Luma AI, InVideo AI, VEED AI, and Leonardo.ai all bundle credit pools into subscription tiers that meter each generation. Pure transcription and infra APIs — Deepgram, Gladia, Daily, LiveKit — tend to quote the raw per-minute (or per-hour) rate without a credit layer. The further from a developer API and the closer to a creative tool, the more likely the minute is wrapped in prepaid credits.

• Free minutes are the standard on-ramp, with enterprise holdouts. Daily gives 10,000 free participant-minutes/month, Gladia grants 10 free hours/month, Twelve Labs seeds a one-time 600 indexing minutes, and Retell AI starts every account with $10 in credits. The exceptions are sales-led enterprise voice: PolyAI and Hippocratic AI publish no self-serve tier at all, betting that a health system or contact center buying agent-minutes at scale is past the trial stage.

Counterexamples & variants

The most common variant is the vendor that generates speech but bills it by the character, not the minute. Deepgram’s Aura TTS is priced per 1,000 characters, Speechmatics TTS is per 1,000 characters, and Murf AI’s API is $0.01–$0.03 per 1,000 characters. These companies meter media minutes for transcription or agents but switch to per-character billing for synthesis, because the input to synthesis is text of known length, where the input to transcription is audio of unknown word count. The same vendor runs two units side by side, and only one of them is the minute.

PolyAI is the variant that proves the minute can be the meter without ever appearing on a price list. PolyAI publishes no public pricing and sells enterprise contact-center voice as a sales-led annual contract billed per-minute (varying by STT and voice configuration) plus one-time voice-design fees. Parloa works the same way — a /pricing path that 404s and a sales-led enterprise deal underneath. The connected minute is almost certainly the underlying cost driver, but the buyer negotiates a sales-led contract rather than reading a per-minute rate. Hippocratic AI is the same idea reframed as labor: ~$9 per agent-hour of active patient time, positioned as on-demand “AI staffing” you hire like a nursing shift rather than an API you meter.

WellSaid Labs sits at the opposite extreme: it produces per-minute media (AI voiceover) but bills largely by the seat rather than by minutes consumed. For a content team that generates voiceovers all day, a flat seat removes the per-minute anxiety entirely. Otter.ai and Fireflies.ai lean the same way — per-seat freemium subscriptions where transcription minutes are an allowance cap inside the plan (Otter meters a monthly transcription-minute limit per tier) rather than a metered line item. Descript and Krisp also lead with per-seat subscriptions and treat minutes as a bundled entitlement. In these cases media minutes exist in the taxonomy but are a usage grant, not the unit the buyer transacts in.

Finally, Kustomer and Fal show the minute as a secondary meter. Kustomer is a seat-priced CRM whose Voice channel is a pay-as-you-go add-on from $0.02/minute — the minute rides alongside seats and per-resolution AI charges, not as the headline. Fal is a generative-media GPU platform that bills most models per output (per image, per video) but exposes per-second video rates that resolve to a media-minute unit. Even LiveKit stacks the minute as one meter among several — agent-session minutes, WebRTC media minutes, inference credits, telephony, and bandwidth all bill in parallel. The minute appears, but as one line among several rather than the spine of the pricing.

What this means for buyers vs vendors

For buyers

Estimate your monthly minutes before you compare rates — your bill is dominated by volume, not by the headline number. A team running 100,000 agent-minutes/month sees a real difference between Retell AI at ~$0.115/min composed cost ($11,500) and a cheaper bundled quote, but a team doing 500 minutes/month will barely feel it. Match the meter to the task: transcription is metered per minute of input audio, synthesis is usually per character of input text, and voice-agent minutes are an unbundled stack — so a “voice AI” quote needs to be split into infra, LLM, TTS, and telephony before you can compare it. Check the metering base — Daily bills per participant-minute (a 4-way call costs 4x), Resemble AI bills per second, and Gladia quotes per hour, so normalize everything to a single unit before comparing. Watch for the credit layer: when Synthesia, Opus Clip, or Luma AI sells you credits, convert them back to minutes (Opus Clip is roughly one credit per minute of video) so you are comparing minutes to minutes. And if you generate media all day, price the seat-based variant — Otter.ai, Fireflies.ai, and WellSaid Labs may beat any per-minute meter for high-volume teams. Use free minutes to run a real pilot: Daily’s 10,000 free minutes and Gladia’s 10 free hours are enough to load-test before you commit. See choosing the right usage metric and the introduction to usage-based pricing for the framing.

For vendors

The media minute is the most intuitive meter you can offer a speech or video buyer — they already think in call length and video duration — but it is also the most directly comparable, so your per-minute rate sits next to every competitor’s. Differentiate inside the minute rather than on it: split accuracy or speed tiers the way Twelve Labs separates Marengo indexing from Pegasus analysis, ladder a model range like Hume AI’s $0.02–$0.072/min empathic voice, or expose the unbundled stack transparently like Retell AI and Vapi so buyers self-select the rate that matches their need. Decide deliberately whether to expose the minute or wrap it: a developer API wins on a transparent per-minute (or per-hour) card (Daily, Gladia), while a creative tool can escape rate-card comparison by selling credits that convert to minutes (Synthesia, Pika, Luma AI). Use a free-minute allowance as the on-ramp — Daily’s 10,000 free minutes/month is a low-friction trial — unless your buyer is a sales-led enterprise past the experimentation stage, as PolyAI and Hippocratic AI bet. Whatever you choose, you need per-second attribution of media duration to a customer and a job — and for voice agents, attribution of every component of the composed minute — which is a heavier metering pipeline than counting requests; see tracking and metering usage events and billing cycles and invoicing.

Explore this theme in the knowledge graph →