I get some version of “how much will this actually cost me” in my inbox every week from readers building podcast workflows, call-center dashboards, or YouTube pipelines around AI. The honest answer is that “AI audio pricing” isn’t one number — it’s three completely different products wearing the same marketing language, and mixing them up is how people blow their budget in month one.

How I Tested and Priced This

I run websites2know.com, where I test AI and SaaS tools the way a non-developer buyer actually uses them — signing up, hitting a pricing page, running a real file through the product, and seeing what shows up on the invoice. For this piece I did two things.

First, I pulled the September 2026 list prices directly from each provider’s own pricing page or model documentation for nine developer-facing models and five consumer apps, then converted every rate to a single unit — cost per hour of audio or video — so they could actually be compared. Second, I opened accounts on Rev, Otter.ai, Descript, and Sonix, uploaded the same 47-minute recorded podcast interview to each, and timed how long it took, what the transcript looked like, and where the pricing page quietly disagreed with the checkout screen.

A few things surprised me. Descript’s checkout no longer shows a per-hour rate at all — it bills in “media minutes” and separately metered “AI credits,” and I had to run a second file through just to figure out how many media minutes my 47 minutes actually consumed. Otter’s free plan gives you three file imports for the lifetime of the account, not three a month, which I only discovered after my third upload silently failed. And Rev’s AI tier undercut its own marketing copy on accuracy for a noisy segment of my interview recorded over a car’s Bluetooth speaker — solid opening paragraph aside, it needed a manual cleanup pass I hadn’t budgeted time for.

Everything below carries that as-of date. Two of these prices are scheduled to change on a specific day, and I’ve flagged both.

Three Products, Three Price Tags — Don’t Compare Them Directly

“AI audio processing” gets used as a catch-all, but you’re actually choosing between three different jobs.

Speech-to-text transcription turns audio into a verbatim written transcript. It’s billed per minute or per hour of audio, and the transcript itself is included in that price. This is what Rev, Otter, Deepgram, and OpenAI’s Whisper endpoint all do.

Audio understanding means handing the recording to a multimodal model like Gemini or GPT-audio and asking it a question — “summarize this call,” “who committed to what deadline.” You’re billed per audio token for the input, and whatever the model writes back is billed separately as output tokens. There’s no transcript unless you specifically ask for one, and asking for one drives the output cost up fast.

Video understanding is the same idea with frames added. The model samples the video at a set rate — usually once per second — and price scales almost entirely with the resolution you feed it, not the length of the clip. This is the one most buyers get wrong, because a “cheap” video model at low resolution can cost 3 to 8 times more once you bump it to high resolution for the same footage.

Developer API Pricing: Cost Per Hour of Transcription

These are the raw, pay-as-you-go API rates you’d use if you or a developer are building your own pipeline rather than using a finished app. Batch (pre-recorded) is separate from live/streaming, because the two are priced on different curves.

Batch, pay-as-you-go, standard tier. List prices read from each provider’s pricing page or model card, September 2026.

Live transcription is priced on a separate, steeper curve because streaming imposes latency requirements the batch endpoints don’t have to meet.

If you don’t need a live caption or real-time agent transcript, batch is almost always the cheaper — and honestly the more accurate — choice.

What It Actually Costs With No-Code Tools (Rev, Otter, Descript, Sonix)

Most of my readers aren’t going to write a script against the Deepgram API. They want to know what it costs to open an app, drop in a file, and get a transcript back — and that number is a lot higher than the developer rates above, because you’re also paying for the interface, the editor, and the support.

Effective cost per hour assumes the plan’s minute allowance is fully used; going over it resets you to overage or top-up pricing.

My honest opinion after testing all four: Sonix is the best pure value if all you want is a transcript — $5 an hour with 53+ languages beats Rev’s AI tier by a wide margin, and I found its editor less fiddly than Descript’s for a straight transcription pass.

Otter is the right call specifically for live meetings, not uploaded files — its Business tier’s unlimited meeting transcription only applies to Zoom, Meet, and Teams bots, and imported files still burn against a monthly minute cap that’s easy to blow through if you also process recorded interviews.

Descript is genuinely the best tool here if you’re editing a podcast or video and want to cut by deleting words from a transcript. It is not the tool to reach for if transcription is the whole job — you’re paying for an editor you won’t fully use, and the media-minutes-plus-AI-credits system makes your monthly bill harder to predict than a flat per-hour rate.

Audio Understanding: When Listening Beats Transcribing

If you don’t need every word written down — you need “what did this call agree to” or “summarize this podcast episode” — sending the audio straight to a multimodal model’s input can be an order of magnitude cheaper than transcribing it first. Here’s the input-only cost per hour; the model’s answer is billed separately as output tokens.

OpenAI doesn’t publish a per-second audio-token rate for these models directly; the figures use OpenAI’s own documented Realtime API rate as an estimate, which is why the pricing page and my table won’t match exactly.

The catch nobody puts in the headline: the model still has to write its answer, and a full verbatim transcript as output would cost more than the audio input did. This route wins big for summaries and action items, not for “give me every word back.”

Video Understanding: Why Resolution Is the Real Price Tag

Video pricing looks confusing until you realize it’s not really priced by length — it’s priced by how many pixels the model looks at every second. The same hour of footage into the same model can cost 3 to 8 times more just by switching a resolution setting from low to high.

One hour of video with its audio track, one frame per second, input tokens only. A single unchunked hour into Gemini 3.1 Pro crosses a 200,000-token threshold that doubles the rate — chunking the video avoids that.

What This Costs in Real Workflows

List prices don’t mean much until you attach them to an actual monthly workload. Here’s how four common jobs shake out.

Solo podcaster, 4 hours of raw audio a week (~17 hrs/month): Whisper API on your own script runs about $6/month. Rev’s pay-as-you-go AI tier runs about $255/month for the same volume. Otter Pro’s 20-hour cap fits with room to spare for $8–17/month.

Support call center, 500 hours/month: Deepgram Nova-3 lands around $129/month; AssemblyAI Universal-2 around $75/month. Rev’s consumer AI tier at $15/hour would run $7,500/month for the same volume — a call center should never be pricing off the consumer page.

YouTuber repurposing 10 hours of raw footage/month into clips and captions: Gemini 3.8 Flash video input at low resolution runs about $2.75/month before output tokens for the analysis pass; Descript’s Creator plan at $24–35/month bundles the actual editing on top.

Grad student transcribing 20 hours of interviews for a thesis: AssemblyAI’s raw API cost is around $3 total. Otter Pro’s 1,200-minute cap covers it comfortably for $8.33–16.99/month, which is the more realistic option if you don’t want to touch an API key.

How to Estimate Your Own Monthly Cost

1. Total your monthly audio/video hours. Be honest — count what you’ll actually process, not your best-case estimate.
2. Decide which of the three products you need. Verbatim transcript, a summary/understanding, or a video pass. This alone can swing your cost by 10x.
3. Pick 2–3 candidates from the tables above that fit your product choice and check whether their free tier or a $5 test run covers a sample of your real content.
4. Run one representative hour through each candidate and read the actual token counts or minutes billed — don’t trust the marketing page alone. This is exactly where Descript’s media-minute system and OpenAI’s estimated per-minute rates for token-billed models can surprise you.
5. Multiply by your monthly hours, add any output/writing cost for understanding models, and compare that number against a subscription tier’s flat price — sometimes the flat plan wins even if the per-hour math looks worse on paper.
6. Re-check pricing every quarter. Two rates in this article change on fixed dates, and providers revise pricing pages without much notice.

My Picks by Use Case

Costing out an AI pipeline is the same exercise whether it’s audio, video, or writing SEO content with AI — you’re always trading a cheaper, sloppier model against a pricier, more reliable one. I’ve seen the same trade-off play out when comparing Mistral’s Voxtral models against the bigger names, and again when I looked at whether ChatGPT Pro is worth $200 a month for a small business.

If your audio actually comes from phone calls rather than recordings, it’s worth reading my 3CX review before you build a separate transcription pipeline on top of a phone system that may already offer one, and the same logic applies if your source is video meetings rather than uploaded files.

For creators pulling captions and clips out of raw footage, I’d also point you to my rundown of UGC video generation tools, and if you’re evaluating a newer open-weight model against these established providers, my Kimi K3 first-impressions review covers similar ground on cost versus capability. Anyone pricing several AI subscriptions at once may also find my breakdown of Geekflare’s credit system useful for spotting the same kind of hidden-cost pattern.

FAQ

Why is transcribing audio more expensive than feeding it to an AI model?

Transcription bundles the output — the written transcript — into the price. Audio understanding charges only for reading the audio in; if you also ask the model to write out a full transcript as its answer, the output tokens usually make it more expensive than dedicated transcription, not less.

Is Whisper still worth using in 2026?

At $0.36/hour it’s no longer the cheapest option — Deepgram, AssemblyAI, and several newer entrants beat it on price — but it remains reliable, well-documented, and easy to self-host for free if you have the compute, which none of the hosted competitors offer.

Why does video cost so much more than audio?

Because you’re paying for frames, not just seconds. A model sampling video at one frame per second and high resolution can consume ten times the tokens of pure audio input for the same clip — resolution and frame rate, not duration, drive video cost.

Should I use a subscription app or pay-as-you-go API pricing?

If your volume is steady and under a plan’s minute cap, a subscription like Otter Pro is usually cheaper and far less hassle. If your volume spikes or you need to build the output into your own product, API pricing scales better and avoids paying for unused minutes.

Are these prices likely to change again soon?

Almost certainly. Gemini 3.8 Flash’s input price is scheduled to double on January 1, 2027, Microsoft’s preview transcription rate holds only through December 31, 2026, and Deepgram’s live-transcription rate is marked as a limited-time promotion. Treat every number here as a snapshot, not a guarantee.

What’s the cheapest way to caption a YouTube video?

A batch transcription API like AssemblyAI or Deepgram, run on the raw audio track, is cheaper than any consumer app for pure captioning — you’re looking at roughly 15–27 cents per hour of footage before you touch an editor.

Conclusion

An hour of AI audio or video processing costs anywhere from under a cent to several dollars, and the gap has almost nothing to do with quality — it comes down to which of the three products you’re actually buying and, for video, which resolution setting you leave switched on.

If I had to leave you with one rule from testing all of this myself: price the job, not the provider. A transcription API, a listening model, and a subscription app are not interchangeable line items, and the cheapest number on a pricing page is only the right number if it’s solving the problem you actually have.

I’ll keep this page updated as the dated prices above shift — bookmark it rather than a screenshot of today’s numbers.