Whisper API Pricing: What the Text Step Adds

Cost per job on the text side

Speech-to-text and text generation are priced in different units, so it helps to look at them separately. This page covers the text half, because that is the part this service sells. The numbers below come straight from the price list: $0.15 per million input tokens and $0.50 per million output tokens.

Job	Input tokens	Output tokens	Cost
10-minute standup, short summary	2,000	300	$0.00045
1-hour meeting, summary and action items	12,000	800	$0.0022
3-hour workshop, structured notes	36,000	1,500	$0.00615
Near-full window, long report	96,000	4,000	$0.0164

Token counts here are estimates. A rough rule for English speech is that one hour of talking produces about 9,000 words, which is on the order of 12,000 tokens, but real numbers depend on the speaker and on how the transcript is formatted.

How the math works

Take the one-hour meeting. Input is 12,000 tokens, so 12,000 / 1,000,000 x $0.15 = $0.0018. Output is 800 tokens, so 800 / 1,000,000 x $0.50 = $0.0004. The total is $0.0018 + $0.0004 = $0.0022, a bit over two tenths of a cent.

Output tokens cost more than three times as much as input tokens, which sets the strategy: reading a long transcript is cheap, writing a long answer is where the money goes. A summary that is 5 percent of the transcript length costs far less than a rewrite of the entire transcript.

The same calculation as a function, in case you want to run it against your own volumes:

Copy# Base URL https://api.whisperapis.com/v1, model "uncensored"
def text_cost(input_tokens, output_tokens):
    # $0.15 per 1M input tokens, $0.50 per 1M output tokens
    return input_tokens * 0.15 / 1_000_000 + output_tokens * 0.50 / 1_000_000

print(round(text_cost(12_000, 800), 6))    # one-hour meeting -> 0.0022
print(round(text_cost(36_000, 1_500), 6))  # three-hour workshop -> 0.00615

Check the arithmetic on the largest row as well: 96,000 x $0.15 / 1M = $0.0144, and 4,000 x $0.50 / 1M = $0.002, which adds up to $0.0164.

The audio side is a separate line item

This service does not transcribe audio, so it has no per-minute audio price. You will need a Whisper-based service of your own choosing for that. Hosted transcription from various providers generally bills by the minute or hour of audio, while a self-hosted setup replaces that fee with the cost of a GPU or CPU time, plus your own maintenance. Prices change often, so check their current pricing instead of relying on a number copied from a blog.

For budgeting, add the two sides together: audio hours times the speech rate, plus transcripts times the token rate above. In most pipelines that process meetings, the transcription line is the larger of the two, because the text step on a one-hour call costs a fraction of a cent. If you run Whisper yourself on hardware you already own, the text API may be the only direct bill you see.

If you want the setup details, the quickstart shows the request format for the text step.

Prepaid credit, no free tier

There is no free tier and no subscription. You pay from prepaid credit, and credit does not expire. New users get one way to test without paying: after sign-up, the API key page has a one-click button for $0.50 of trial credit. It is one per person, valid for 7 days, and needs no card.

That $0.50 goes further than it sounds. At $0.0022 per hour-long meeting, it covers about 227 summaries (0.50 / 0.0022 = 227.3). The trial key is limited to 2 concurrent requests and 60 requests per minute, which is plenty for testing a pipeline on a batch of recordings.

After the trial, you top up with any whole amount from $20 to $500. You pay by card (Visa, Mastercard, Apple Pay, Google Pay) or crypto (USDT on TRC20, USDC on Base). Larger top-ups get a bonus on the credit you receive:

Top-up	Bonus	Credit received
$20	none	$20.00
$50	+5%	$52.50
$100	+10%	$110.00
$200	+15%	$230.00

The pricing page has the current list in one place.

What a month of meetings costs

Suppose a team records 40 one-hour meetings a week. That is about 173 a month (40 x 52 / 12 = 173.3). Each costs $0.0022 on the text side, so the monthly total is roughly 173.3 x $0.0022 = $0.38. A $20 top-up would cover more than four years of that workload, and the credit has no expiry date.

Now a heavier case: a service that processes 5,000 hour-long recordings in a month with a 12,000-token input and an 800-token summary each. That is 5,000 x $0.0022 = $11. Raise the output to 3,000 tokens per recording because customers want detailed notes, and the output cost per recording becomes 3,000 x $0.50 / 1M = $0.0015, so the per-recording total is $0.0018 + $0.0015 = $0.0033, and the month comes to $16.50.

Notice what moved the bill. The transcript length did not change, only the answer length. If costs grow faster than expected, look at your output size before anything else.

Comparing with doing it by hand

The honest comparison is against your own time. An assistant who listens to an hour-long recording and types up notes spends an hour or more. The model reads the same hour as text in a few seconds and writes the notes in perhaps half a minute. At $0.0022 per hour of meeting, the text step is not where the money goes; the audio transcription, the engineering time to build the pipeline and the review of the output cost more.

The model also makes mistakes, so budget a short human check on anything that matters. A reviewer who skims generated notes for two minutes is still faster than writing them from scratch, and the check catches the occasional invented owner or misheard figure. Failed or refused requests cost nothing, so the only text spend is on answers you actually received.

Ways to keep the text bill small

• Cap the answer. Set max_tokens when you only need a short summary. It is optional, but it prevents a runaway answer on an odd transcript.
• Do not resend the transcript. If you need three outputs from one meeting, ask for them in one request as a JSON object instead of three calls that each pay for 12,000 input tokens.
• Read usage from the response. The usage block reports prompt and completion tokens, and streamed responses include usage at the end. The X-Balance-USD response header shows your remaining credit.
• Keep retries cheap. Failed or refused requests are not charged, so a retry loop on errors does not add to the bill.

Copyfrom openai import OpenAI

client = OpenAI(base_url="https://api.whisperapis.com/v1", api_key="YOUR_KEY")
messages = [{"role": "user", "content": "Summarize in five bullets: ..."}]

resp = client.chat.completions.create(
    model="uncensored",
    messages=messages,
    max_tokens=900,
)
u = resp.usage
print(u.prompt_tokens, u.completion_tokens)
print(text_cost(u.prompt_tokens, u.completion_tokens))

Pair those habits with a daily check of the balance header and you will see a cost problem within a day, not at the end of the month. The docs list the response headers and error codes.

More guides

• Whisper API Server: Setup Options and the Text Step
• LLM API Meeting Notes: Transcript to Summary in Python
• Midjourney API: Troubleshooting & Alternatives