Whisper cost per audio hour: rates converted, crossovers worked

Whisper costs between 0.012 and 0.36 per audio hour across the hosted rates read for this page on 26 October 2026. Below roughly 30 audio hours a month, the hosted side wins on every tier worked here. Past roughly 7,400 hours a month, an owned card wins against every rate read here, budget floor included. Between those bounds the answer is a table lookup. This page is the tables: every named rate converted to per audio hour, local power priced per audio hour at three speed tiers, and the crossover worked for each party against each card tier.

All amounts are United States dollar amounts written as bare numerals. This page has no calculator widget. Every table is static and every input is named and dated. You can check every number with a pencil. The math is shown before the tables that use it.

Last verified: 2026-10-26

The answer in two bounds

The whole question is volume. Two formulas carry this page. The first turns any printed rate into the unit that matters, the audio hour:

per audio hour  =  per minute x 60
per minute      =  per hour / 60

The second finds the crossover. That is the audio-hour count where a card you own beats a rate you pay. The full formula, with every variable defined, lives in the platform's Whisper break-even guide. This page uses its short form:

crossover hours  =  G / (P - m)

G:  what the card cost you
P:  hosted rate per audio hour
m:  your power per audio hour
monthly volume   =  crossover hours / 24

The 24 is the payback window in months. The guide's default is twenty-four months, and these tables keep it. Work the easy case end to end. A used 12 GB card at 279, the top list rate of 0.36 per audio hour, power at 0.0031 per audio hour. The crossover is 279 divided by 0.3569. That is about 780 total audio hours, or roughly 33 hours a month. That is the fastest payback anywhere on this page. Below that volume, no card in these tables pays back against any rate here.

The hard case sits at the other end. The budget floor print is 0.012 per audio hour. Against a used 24 GB card at about 1,000, the crossover is about 178,000 hours, roughly 7,400 a month. This pair of cases is the honest headline. Volume below roughly 30 audio hours a month: pay per audio hour. Volume past roughly 7,400: own the card, even against the floor. Between them, use the tables.

Every hosted rate, converted to per audio hour

Every row below was read from the named party's own pricing page on 26 October 2026. Per-minute rows convert by multiplying by 60. Per-hour rows convert by dividing by 60. Some pages print both forms. Where both print, they agree.

Sources, read first-hand: OpenAI pricing, Deepgram pricing, AssemblyAI pricing, ElevenLabs API pricing, the DeepInfra model page. The AssemblyAI sheet carries its own date, 2026-09-28. The Groq cell is dashed on purpose. Its public pricing sits behind a signed-in console today, and this page prints no rate it has not read first-hand. Two secondhand reads put the turbo rate near four hundredths of a dollar per audio hour, dated August and September 2026. A dashed cell is data too.

The spread comes first. A factor of thirty separates the floor print from the top list rate. That factor, not any single number, is why a cost question in audio hours needs tables.

Three billing details move real bills. Deepgram prints true per-second billing. Two-channel audio bills per channel at AssemblyAI, so a two-channel file counts as two hours. Every party here bills the duration of the audio itself, not the time it takes to process.

The local side: power per audio hour

Local Whisper's marginal cost is power. The math is three factors:

power per audio hour  =  kW x price x RTF

kW:    card draw at the wall, kilowatts
price: power price per kilowatt-hour, 18.2 cents national average
RTF:   real-time factor, GPU seconds per second of audio

The power anchor is the government read. The EIA Short-Term Energy Outlook puts the 2026 United States home average at 18.2 cents per kilowatt-hour. The October 2026 edition still carries it. Your own bill is the number that matters. 18.2 is the anchor.

The speed tiers are the pinned set from the break-even guide: RTF of 0.05, 0.10, and 0.15. The guide labels them assumed until a measured figure replaces them. The draws are the makers' own specs: 170 W for a 3060-class card, 350 W for a 3090-class card, and 450 W total graphics power for a 4090-class card per NVIDIA's spec page.

The guide's range for the 350 W box is 0.003 to 0.010 per audio hour across the same three tiers. The second row shows it to four decimal places. Read it plainly. Local power is three tenths of a cent to a full cent per audio hour on the 24 GB card. Paying off the card, not power, is the real local cost. The guide draws the same conclusion.

Community speed reads today beat the pinned tiers. An April 2026 batched test on a 3090-class card reports a full hour of audio in under a minute. A 2026 benchmark roundup puts a 4090 at 35 to 49 times real time. Faster tiers only shrink the power term. The pinned tiers are the careful end of the local side.

Card prices move, so both anchors carry dates. The tables use the guide's frozen anchors: 279 for a used 12 GB card and about 1,000 for a used 24 GB card, September 2026 reads. Today's second-hand reads, October 2026: roughly 180 to 250 sold for the 12 GB card, roughly 950 to 1,100 for the 24 GB card. The formula does not care which read you trust. Put your own number in G.

The crossover tables: when a bought card wins

Rates for this platform's own audio lanes render below from the live catalog. The crossover table reads its P for the platform row off that table. Every other P comes from the first-hand reads above.

Both tables below run at the 0.10 speed tier and the twenty-four-month window. Power comes from the local table: 0.0031 per audio hour on the 12 GB card, 0.0064 on the 24 GB card. Card anchors are 279 and 1,000. One row worked in full, the top rate. On the 12 GB card: 279 / (0.36 - 0.0031) = 780 total hours, about 33 a month. On the 24 GB card: 1,000 / (0.36 - 0.0064) = 2,800 total hours, about 118 a month.

The platform row quotes the published break-even guide. The guide works this platform's planning class, a hosted rate near a tenth of a dollar per audio hour. Its shapes: 2,500 to 2,900 hours on the 12 GB card, 9,000 to 11,000 on the 24 GB card. The live rate in the fence sits in that class today. It is about a third of the top list rate, the same relation the guide records. Read the exact figure off the fence, never off this prose.

One caveat belongs on the floor row. The 0.012 print is a card price on a model page, and a print is a claim. In September 2026, one budget host's meter billed about one hundred times its own printed rate. Measured audio and a wallet check caught it. The break-even guide records the case in its drift stories. Treat any floor print, this one included, as unverified until a measured minute says otherwise. If the floor's true metered rate is higher, its crossover shrinks toward the mid-band rows.

One rule falls out of the shape. The denominator is the gap between the rate and your power. At floor rates the power term eats most of the gap, and the crossover explodes by orders of magnitude. At mid and top rates the power term is noise, and the crossover is card cost divided by price. The full formula, with the payback term live, is the break-even guide's to own.

The rented middle, priced honestly

Every ranking page that sells you a middle option prices the extremes wrong to do it. The honest version prices the middle too. Two rental classes, read for this page:

A rental is a subscription wearing ownership's clothes. It never becomes hardware you own. It bills around the clock whether audio arrives or not. The rental pages' own crossovers, 300 to 550 hours a month, all compare rent against the top rate only. Run the same rentals against the floor print and the mid band and the picture changes. At 0.012 per audio hour, 730 card-hours of the entry tier costs about 66. The same month's audio at 100 hours costs 1.20 hosted. Renting to save money only works when the card runs packed. That is a scheduling project, not a line item.

At the street prices above, months of rent buy the card outright. That is the frame check the break-even guide runs on the vendor pages. It still holds at today's reads.

The meter is the price

A rate table, this one included, is a set of claims. The invoice is the fact. The check costs less than a coffee. Send one measured minute to the route you are pricing. Then read the wallet. Scale the result to your month. If the meter and the card disagree, trust the meter, and re-run the crossover with the metered number in P.

Three habits fall out of the check. Read rates from the party's own page, dated. Never take a rate from a summary panel or an aggregator. Dash what you cannot read, as this page dashes one cell rather than guess it. And re-read the two or three rates that matter to you on the day a decision rides on them. Rates moved in 2026 in ways that moved answers by factors, not decimals. The guide's drift stories record them.

This platform's own metering is built for that check. Audio lanes bill by the audio second, taken from the duration the transcript itself reports, never from upload wall-clock. A request refused before work starts bills nothing. Spending draws down a prepaid wallet, so the cap applies to arriving work, not to a bill at month end. The check needs no special access: a small wallet, a key, one file.

This platform's whisper lane

The whisper route on this platform is live as of 26 October 2026. It takes the OpenAI transcription request shape. Existing client libraries migrate with a base URL and a model id, and the whisper model page carries the full request contract. Its rate renders from the live catalog in the fence above and on the pricing page. This page never types it. The lane sits mid-band in the ladder. It is below every top-tier read above and above the unverified floor print. That is the position the guide recorded for it.

The audio lane family: hayamimi is the Japanese-first specialist with its own measured comparison. The speech-to-text category page carries the current state of every audio route. The speech-synthesis side of the same math, its own crossover and formula, is the TTS break-even guide.

Work it with your own numbers

Four inputs, two formulas, no code.

1. Convert your rate: per minute times 60, or per hour divided by 60.
2. Price your power: kilowatts times 18.2 cents times your speed factor, or take a row from the local table.
3. Find your crossover: card cost divided by (rate minus power), then divide by 24 for the monthly figure.
4. Check the meter before you commit: one measured minute in, wallet out.

Rates for every listed lane render on the pricing page. To run the check on this platform's whisper route: sign up, create a key, send one file, and read the duration the transcript reports. That number, times the fence rate, is what the next thousand hours will cost.

Frequently asked questions

How much does one audio hour of Whisper cost?

Across the first-hand reads dated 26 October 2026: 0.36 at the top list rate. 0.26 to 0.29 across Deepgram pre-recorded tiers. 0.15 to 0.21 at AssemblyAI. 0.22 at ElevenLabs Scribe. 0.18 on OpenAI's own mini model. 0.012 at the budget floor print. The factor of thirty between floor and top is the real answer. There is no single price. That is why the conversion table and the crossover tables exist.

Is Whisper billed by the minute or by the second?

Rates print per minute or per hour, and rounding differs by party. Deepgram prints true per-second billing on its own sheet. Others round within the minute. Plan on paying for the full duration you send, silence included, and check the party's own billing note before you commit. On two-channel files at AssemblyAI, each channel counts.

When is running Whisper myself cheaper?

Below roughly 30 audio hours a month, never: every hosted rate on this page beats a bought card on both card tiers. Past roughly 7,400 audio hours a month, always: even the 24 GB card pays back against the budget floor print. Between those bounds the party and the card decide, and the crossover table gives the exact pair. The floor row carries a caveat. Its print is unverified against its meter, and a higher metered rate moves its crossover toward the mid-band rows.

Does electricity decide anything?

No. Local power is 0.003 to 0.010 per audio hour across the pinned speed tiers on a 24 GB card. That is three tenths of a cent to a full cent. Against rates of 0.15 to 0.36, it is noise. It only starts to matter at the budget floor, where the power term eats most of the gap between the rate and zero. Paying off the card is the local side's real cost, and the break-even guide works it in full.

Why is one rate cell dashed?

Because its owner's pricing page sat behind a signed-in console when this page was written. A rate is printed here only when it was read first-hand from the party's own page on the date shown. Secondhand reads put it near four hundredths of a dollar per audio hour, but secondhand is secondhand. The dash is the honest cell. The same rule is why the platform's own rate renders from the catalog fence instead of appearing as a typed number.

The two bounds travel well. Under 30 audio hours a month, pay per audio hour. Past 7,400, own the card. Between them, the tables. Every rate above was read 26 October 2026 from the named party's own page. The platform's own rates live on the pricing page, rendered from the same catalog the meter reads. To price your own volume the honest way: sign up, fund a small wallet, and measure one minute before you measure a month.

Last verified: 2026-10-26