Transcription costs in 2026: every published rate on one axis

Put every 2026 transcription price on one axis – dollars per audio minute – and the spread is roughly 300x. At one end, a raw cloud API bills a fraction of a cent. At the other, professional human work runs a dollar or two a minute, and court reporting is billed per page on a government schedule. Same job, three orders of magnitude apart.

The tired comparison is "AI is cheaper than humans." Everyone knows that. The useful question is why each tier costs what it costs, because once you see the reason you know which one you actually need. Here is the whole range, cheapest first. Every price below carries a "Last verified: 2026-07-12" note, because vendor rates change without notice and the federal court schedules update roughly once a year. For a plainer walkthrough of the same ground, see our guide to how much transcription costs.

The cheapest number anywhere is the raw API. OpenAI bills gpt-4o-transcribe at $0.006 per minute ($0.36 an hour) and gpt-4o-mini-transcribe at $0.003 per minute ($0.18 an hour). Last verified: 2026-07-12.

There is no mystery to the price. You send audio to an endpoint and get text back. No editor, no storage, no account to log into, no support queue, and above all no human in the loop. This is the floor everything else is built on top of. It is also raw material, not a product: you get a transcript, but chunking long files, stitching timestamps, labelling speakers, and formatting exports are all still yours to build. That trade-off is the whole story of our developer notes.

One step up from the raw API is the packaged AI tool. Rev's automated transcription is $0.25 per audio minute on its pay-per-minute plan – one-eighth of its own human rate. Sonix lists automated transcription at $10 an hour pay-as-you-go (about $0.17 a minute), or on subscription: Core is $25/month for 5 hours, Advanced $50/month for 20 hours, and Pro $80/month for 40 hours. GoTranscript's own guide pegs AI transcription at $0.05-$0.25 a minute and adds the honest caveat that the output still needs proofreading. Last verified: 2026-07-12.

The jump from $0.003 to a quarter is the product: a browser editor, cloud storage, exports, speaker labels, a support inbox. You are renting the convenience of not touching the API yourself. Just remember what the cents buy you – a first draft, not a finished document. How close that draft lands to publishable depends on your audio, and we cover the realistic gap in how accurate AI transcription is.

Now the price jumps two orders of magnitude, because a person is doing the work. Rev's human transcription is $1.99 per audio minute as a base rate (Rev subscribers get 10-15% off human transcription on its Essentials and Pro plans). GoTranscript puts general human transcription at $0.70-$2.00 a minute – roughly $60-$120 per audio hour – and prices specialised legal and medical work higher, at $1.50-$3.50+ a minute, because it demands expert terminology. Last verified: 2026-07-12.

Why is this 300x the API floor? Labor, and the arithmetic of it. Published estimates put manual transcription at roughly 4-6 hours of work per hour of audio, so one hour of clean human transcript is really the better part of someone's day. You are not paying for a GPU's second. You are paying for a person's afternoon, plus the judgment to get names, jargon, and crosstalk right that a first-pass model still fumbles.

Court reporting does not sit on the per-minute axis at all. It is unbundled: a per-page transcript fee plus a daily appearance fee, with optional realtime and video surcharges on top. The U.S. federal courts cap transcript pages on a delivery-speed schedule, effective Oct 1, 2024: $4.40 per original page for ordinary (30-day) delivery, $5.10 for 14-day, $5.85 for expedited (7-day), $7.30 for daily (next-day), and $8.70 for hourly (2-hour) turnaround; first copies run $1.10-$1.45. On top of that, federal contract reporters bill a per-diem appearance fee: $300 a day non-realtime, $325 realtime, and $355 realtime certified, with half-days at half. Last verified: 2026-07-12. (The page schedule quoted is the D.C. federal district's; other districts differ slightly.)

What does that add up to? A full day of testimony delivered next-day runs into four figures for the transcript alone. A full-day deposition typically yields 250-400 pages, and at the $7.30 daily-delivery maximum that computes to roughly $1,825-$2,920 before the reporter even bills the appearance fee. Treat that as illustrative arithmetic on the verified page rate, not a quoted flat rate: the pages-per-day range is an industry estimate. The labor behind the price is real, though – the Bureau of Labor Statistics put the median court-reporter wage at $67,310 in May 2024, with the top 10% over $127,020.

The number to anchor on is not the price. It is what you have to walk away with. If the deliverable is a certified, defensible, or verbatim record – a court transcript, a legal deposition, a medical file – you pay a dollar or more a minute, or per page, because a human's accuracy and accountability are the product. Nobody defends an AI draft in court.

If the deliverable is a fast first draft you will read, quote from, search, or clean up yourself, you pay cents, because the machine gets you most of the way and you finish the rest. Most people, most of the time, need the second thing and buy the first out of habit. Match the tier to the job, not to the fear of getting it wrong.

If you are a developer or running high-volume batch work, the honest answer is the raw API or local Whisper, and no product beats them on price. At $0.003-$0.006 a minute you cannot undercut OpenAI's endpoint, and a local Whisper model on your own machine costs nothing per minute and never uploads the audio – which matters for genuinely sensitive material.

The catch is everything around the transcription: splitting long files, stitching the timestamps back together, speaker labels, retries, exports. That plumbing is real engineering, and it is exactly what our developer notes and the long-audio guide walk through. If you have the engineering time, spend it there and pocket the difference.

Pepys is built for the shape in the middle: irregular, self-serve volume where you want AI-draft prices without renting a subscription you will use twice. Credits are pay-once and never expire, so a burst of interviews in March does not leave you paying a monthly bill through December. That is the whole model on our pricing page, and it is the same wishlist we heard reading 100 Reddit threads about transcription: pay for what you use, keep the minutes you don't.

It is not the cheapest possible option – a raw API or local Whisper wins on pure price – and if you need a certified court record, you need a court reporter, not us. Where those are the right answer, they are the right answer, and we would rather say so than pretend otherwise. For everything between "a fast draft I'll clean up" and "I don't want to build the pipeline myself," that is the room we built for. If the subscription cap is the part that stings, that is the case for transcription without a subscription.