You have a folder full of recordings, a deadline, and one nagging question: what will it cost to turn all that speech into text? The honest answer is "less than a coffee" for most people, and "a real line item" for call centers and media archives. The trick is knowing which OpenAI model you are paying for, how the billing actually works, and when a free or cheaper route makes more sense.
This article breaks down OpenAI speech to text pricing in plain numbers, shows what typical projects cost per month, lists the free options that genuinely work, and compares alternatives you can try right now.
💡 Quick answer: OpenAI's transcription models land at roughly $0.003 to $0.006 per minute of audio. That is about $0.18 to $0.36 per hour. There is no permanent free API tier, but there are several workable free routes, covered below.
What OpenAI Charges Per Minute
OpenAI sells speech to text through its API under a few model names. The sticker price is usually quoted per minute, which makes budgeting simple even though the billing underneath is more nuanced.

The Three Price Points
Here are the rates most pricing trackers and OpenAI's own model pages agree on at the time of writing:
| Model | Approx. cost per minute | Approx. cost per hour | Best for |
|---|
| whisper-1 (original Whisper API) | $0.006 | $0.36 | Legacy workflows, subtitles with timestamps |
| gpt-4o-transcribe | $0.006 | $0.36 | Highest accuracy, accents, noisy audio |
| gpt-4o-mini-transcribe | $0.003 | $0.18 | High volume, clean audio, tight budgets |
Two things stand out. First, the premium GPT-4o model costs the same per minute as the old Whisper endpoint, yet it generally handles accents, background noise and fast speech better. Second, the mini variant is half the price, which matters a lot once you pass a few hundred hours.
You can try both GPT-4o options without writing code: GPT-4o Transcribe and GPT-4o Mini Transcribe are both available on PicassoIA.
⚠️ Heads up: Pricing and model lifecycles change. One tracker I checked listed a newer OpenAI model at about $0.0045 per minute and mentioned a retirement date for the mini model in early 2027. I could not confirm either detail, so verify on OpenAI's own pricing page before you commit to a large batch job.
Why Billing Is Token Based
The per-minute figure is an estimate. Under the hood, the GPT-4o transcription models bill by tokens: audio input tokens when the model listens, plus text output tokens for the transcript it writes. OpenAI converts that into an average per-minute cost so budgeting is easy.
What that means in practice:
- Dense speech costs a bit more. A fast talker produces more output tokens than a slow one.
- Silence is not free. Long pauses and hold music still count as audio input.
- Prompts add input tokens. If you pass a long text prompt to steer vocabulary, that text is billed too.
For planning, using $0.006 per minute for the premium model and $0.003 for mini will keep you within a few percent of the real invoice on normal conversational audio.

Real Costs for Real Projects
Rates only mean something when you multiply them by your workload. These examples use the two GPT-4o prices from the table above.
| Audio volume | gpt-4o-mini-transcribe ($0.003/min) | gpt-4o-transcribe ($0.006/min) |
|---|
| 10 minutes | $0.03 | $0.06 |
| 1 hour | $0.18 | $0.36 |
| 10 hours | $1.80 | $3.60 |
| 100 hours | $18 | $36 |
| 1,000 hours | $180 | $360 |
Podcasts and Interviews
A weekly podcast with four 60-minute episodes produces 240 minutes of audio per month. That is $0.72 on mini or $1.44 on the premium model. Even a daily show lands under $10 a month.
Journalists and researchers are in a similar range. An investigative reporter with 30 hours of interview tape pays roughly $5 to $11 to get a searchable transcript. The money is not the issue here. Accuracy on names, jargon and overlapping speakers is, which is why the premium model usually wins for interviews.

Lectures and Meetings
A university department recording 20 lectures a week at 90 minutes each handles 30 hours weekly, or around 130 hours a month. That costs about $23 on mini and $47 on the premium model. Compare that with human transcription, which commonly runs from about a dollar to several dollars per audio minute, and the gap is hard to ignore.

Business meetings follow the same math. A team of 25 people with 10 hours of recorded calls each per month generates 250 hours, which costs $45 on mini or $90 on premium. If those transcripts feed meeting summaries or action item extraction, the speech to text bill is usually the smallest piece of the pipeline.
Call Centers at Scale
This is where the decimals start to matter. Take 50 agents, 6 hours of live talk time per day, 22 working days a month: about 6,600 hours of audio.
- On mini: roughly $1,188 per month
- On the premium model: roughly $2,376 per month
At that volume, choosing mini for clean headset audio and the premium model only for difficult calls (heavy accents, speakerphone, noisy rooms) can save four figures a year without hurting quality where it counts.

Free Options Worth Trying
Is there a free OpenAI speech to text route? Not as a standing API tier. OpenAI generally bills from the first request, though new accounts sometimes receive trial credits, so check your dashboard. Beyond that, you have three realistic free paths.

Open Weights Whisper Locally
OpenAI released the original Whisper model as open weights, so you can run it on your own computer at zero per-minute cost. The trade-offs are real, though:
- You need a decent GPU for reasonable speed. On a CPU, an hour of audio can take longer than an hour to process.
- You maintain the setup yourself: Python, drivers, model sizes, updates.
- Accuracy on the smaller local model sizes is below what the hosted GPT-4o transcription models deliver.
If you transcribe a few files a month and enjoy tinkering, this is great. If you want reliability and zero setup, a hosted model is cheaper than your time.
Free Credits and Trials
Many platforms give new users a small allowance. On PicassoIA you can run speech to text models from the browser with no code, which is handy for testing a model on your own audio before you budget a large batch. Check your current plan for what is included.
💡 Tip: Test on a representative 5-minute clip first: the noisiest, most accent-heavy one you have. If the cheaper model handles it, you have just halved your bill.
Built-In Dictation
If your actual need is dictation (talking to type notes or emails), your phone keyboard, operating system dictation and the voice mode inside ChatGPT already do it without a separate per-minute charge. They are not designed for batch processing recordings or producing timestamped files, but for personal notes they cost nothing extra.
Hidden Costs Behind the Rate
The per-minute price is only one part of the real cost of a transcription workflow.

File Limits and Chunking
OpenAI's transcription endpoint caps uploads at around 25 MB per file. A one-hour MP3 at normal quality can exceed that, and WAV files get there within minutes. You then need to compress or split the audio, which adds scripting work and creates awkward cuts mid-sentence. Splitting on silence rather than at fixed intervals prevents most of the lost words at the boundaries.
Cleanup and Review Time
No speech to text model is perfect. Budget for these:
- Proper nouns and brand names, which get misspelled most often
- Speaker labels, since basic transcription returns plain text, not who said what
- Punctuation and paragraphing for long-form content
- Domain terms in medicine, law or engineering
A short text prompt that lists the unusual names and terms in your recording reduces many of these errors before they happen. That small habit often saves more time than switching models.
Alternatives to OpenAI Transcription
OpenAI is not the only choice, and for some jobs it is not the best one.

Gemini 3 Pro
Gemini 3 Pro from Google is a multimodal model that can listen to audio and return a transcript, and it can also summarize, translate or answer questions about the recording in the same step. That makes it a strong pick for meetings, where the transcript is only the starting point. Use it side by side with GPT-4o Transcribe on your own audio and compare errors on names and numbers.
Other Vendors in Plain Sight
Several dedicated speech vendors compete directly on price and features. Deepgram, AssemblyAI, Google Cloud Speech-to-Text, Amazon Transcribe and Microsoft Azure Speech all offer per-minute or per-hour pricing, often with extras such as speaker labels, live streaming and custom vocabularies. Their rates shift by tier and volume, so compare quotes for your actual workload rather than trusting headline numbers.
Here is a simple way to choose:
| Your situation | Best fit |
|---|
| Lowest cost on clean audio | GPT-4o Mini Transcribe |
| Accents, noise, interviews | GPT-4o Transcribe |
| Transcript plus summary or translation | Gemini 3 Pro |
| Zero cost, you own a GPU | Local open weights Whisper |
| Live captions, speaker labels built in | A dedicated streaming vendor |
How to Use GPT-4o Transcribe on PicassoIA
If you want the premium OpenAI model without managing an API key, billing dashboard or upload scripts, GPT-4o Transcribe on PicassoIA runs entirely in the browser.

Step by Step
- Open the model page. Go to GPT-4o Transcribe.
- Upload your audio file. Accepted formats include MP3, MP4, MPEG, MPGA, M4A, OGG, WAV and WebM. No conversion needed.
- Set the language. Enter the ISO-639-1 code, such as
en for English, es for Spanish or fr for French.
- Add an optional prompt. List names, products or jargon that appear in the recording, or paste the last few lines of the previous segment if you split a long file.
- Leave temperature at 0. For transcription, you want the most literal output.
- Run it and copy the transcript. Paste it into your editor, subtitle tool or notes app.
Parameter Tips
| Parameter | What it does | Recommendation |
|---|
| audio_file | The recording to transcribe (required) | Trim long silences and intros first |
| language | ISO-639-1 code | Always set it: better accuracy and lower latency |
| prompt | Style or continuation text | Match the audio language; keep it short |
| temperature | Randomness from 0 to 1 | Use 0 unless you see repetitive output |
💡 Budget move: Run your first pass on GPT-4o Mini Transcribe. Send only the files with visible errors through the premium model. For most libraries that cuts spending by 30 to 40 percent.
Which Model Should You Pick
Skip the spreadsheet debate and use this short decision path:
- Under 50 hours a month? Cost is irrelevant. Pick the premium model and enjoy fewer corrections.
- Over 500 hours a month with clean audio? Start with mini and sample 5 percent of output for quality checks.
- Mixed audio quality? Route by difficulty: mini for studio and headset recordings, premium for phone calls and rooms with echo.
- Need more than a transcript? Try Gemini 3 Pro and let it summarize and translate in the same pass.
- Privacy rules forbid sending audio to a cloud service? Local open weights Whisper is your only truly offline option.
One more reminder on numbers: the rates here are the averages most sources report right now. Run a 10-minute test, check the usage on your invoice or dashboard, and extrapolate from your own audio rather than from anyone's headline price, including mine.
Try It on Picasso IA Today
Transcription costs fractions of a cent per minute, so the cheapest way to settle the question is to test it yourself. Pull out your noisiest recording, upload it to GPT-4o Transcribe, run the same file through GPT-4o Mini Transcribe and Gemini 3 Pro, and compare the three transcripts line by line.
Picasso IA also brings text to speech, AI music generation, image creation and video tools into the same workspace, so a recording can become a transcript, a voiceover or a visual in a few clicks. Browse the full catalog at picassoia.com/en/all-models and start experimenting with your own audio today.