← All posts

How to Get an OpenAI, Deepgram, or Groq API Key (and What Each Actually Costs)

Short answer - which key should you get?

  • You want to try cloud transcription today, no card → Groq. Free tier, whisper-large-v3-turbo, $0.04 per hour of audio once you upgrade.
  • You want a big free runway → Deepgram. $200 credit, no card, Nova-3 at $0.0043/min for recorded audio.
  • You also want AI summaries from the same key → OpenAI. The only one of the three that does both transcription and text. $5 minimum prepay.
  • You don't want a key at all → you don't need one. Transcription runs on your Mac for free, and summaries can run through a local model.

Full disclosure: I make Humla, a Mac meeting notetaker that takes your own key for any of these three, or none. So I've done this signup dance more times than is healthy, mostly while testing, occasionally while swearing at a billing page. All the prices and steps below come from each provider's own pricing pages and docs, checked 6 October 2026. They move. Check before you budget.

Why you'd want your own key in the first place

Quick context, because "bring your own key" sounds like homework.

Most meeting notetakers bundle transcription into their subscription. You pay them, they pay a speech provider, they keep the margin, and you never see a key. That's fine! It's also why a minute counter shows up on the free tier: every minute you record costs them money.

With your own key, the app is just the app, and the provider bills you directly at cost. Which is suspiciously cheap, as you'll see in a minute. (The first time I added up what a month of my own meetings cost through Groq, I assumed I'd misplaced a decimal. I had not.)

The other reason is control. Your audio goes from your Mac to a provider you picked, under their data terms, which you can read. I've done that reading already, one post per provider: OpenAI, Deepgram, Groq. Worth ten minutes before you send them a client call.

The three at a glance

Card needed to start?Free allowanceTranscription priceAlso does summaries?
GroqNo (free tier)Free tier with rate limits$0.04/hr (turbo)Not in Humla
DeepgramNo$200 credit$0.0043/min Nova-3No
OpenAIYes, $5 minimum prepayNone I could find$0.003-0.006/minYes

Sources: console.groq.com/docs/speech-to-text, console.groq.com/docs/billing-faqs, deepgram.com/pricing, OpenAI pricing, OpenAI prepaid billing. Checked 6 October 2026.

Groq: the no-card one

Getting the key:

  1. Sign up at console.groq.com.
  2. Go to console.groq.com/keys and create a key.
  3. Copy it immediately. It'll start with gsk_, which is what Humla's key field expects.

That's genuinely it. No payment details for the free tier.

What the free tier gets you. Groq's rate limits page lists, for both Whisper models: 20 requests a minute, 2,000 requests a day, 7,200 audio seconds an hour and 28,800 audio seconds a day. So two hours of audio per hour, eight per day. Free-tier files are capped at 25 MB each. For trying cloud transcription on your real meetings for a week, that's plenty.

When you outgrow it. The Developer tier needs a payment method (card, US bank account or SEPA debit), but nothing is charged on upgrade. Groq bills at month end, or earlier when your usage crosses $1, $10, $100 and so on.

What it costs.

ModelPrice per hour of audio
whisper-large-v3-turbo$0.04
whisper-large-v3$0.111

One wrinkle worth knowing: Groq's minimum billed length is 10 seconds per request. Send it a 3-second clip and you pay for 10. Apps that send audio in short pieces as you go (Humla's live chunks are 1 to 15 seconds, cut on pauses) will get some of those rounded up. At four cents an hour, even a pessimistic rounding leaves you paying small change. I'm mentioning it so you're not confused by the invoice, not because it changes the answer.

Humla uses whisper-large-v3-turbo for Groq.

Deepgram: the big free runway

Getting the key:

  1. Sign up at console.deepgram.com. Your first project is created for you.
  2. Open the project, go to Settings → API Keys, and choose Create a New API Key.
  3. Give it a name, optionally an expiry date (handy if you're just testing), and copy the secret. Deepgram's docs are explicit: "we won't be able to show you the key again."

What the free credit gets you. Deepgram's Pay As You Go plan starts with $200 in credit, and the pricing page says "No credit card required." At the Nova-3 rate for recorded audio, $200 is roughly 775 hours. If you have 775 hours of meetings to transcribe before you run out of free credit, I have questions about your calendar, but they're not pricing questions.

What it costs. Humla sends Deepgram finished chunks of recorded audio, so the pre-recorded column is the one that applies.

Nova-3, pre-recordedPay As You GoPer hour
Monolingual$0.0043/min~$0.26
Multilingual$0.0052/min~$0.31

Deepgram bills per second and says it doesn't round up, so a 14-second chunk costs 14 seconds. Streaming is priced separately and higher, and speaker diarization is a paid add-on. Humla doesn't ask Deepgram for speaker labels at all, since it works out who spoke on your Mac after the recording stops, so that add-on never shows up on your bill.

Humla offers nova-3, nova-2 and base for Deepgram.

OpenAI: the one that also writes your summaries

Getting the key:

  1. Sign in at platform.openai.com. This is the API platform, a separate thing from a ChatGPT subscription. Paying for ChatGPT Plus does not give you API credit, which catches a LOT of people.
  2. Go to the billing overview, choose Add payment details, and buy credit. The minimum is $5; the default amount is $10.
  3. Look at the auto-reload toggle before you confirm. OpenAI's help article says it "is turned on by default during setup." If you want a hard ceiling on what you can spend, turn it off.
  4. Create a key at platform.openai.com/api-keys. It'll start with sk-. Copy it straight into wherever it's going.

Two more things from that same help article that I wish I'd known earlier. Purchased credits expire after one year. And running out isn't an instant cutoff: "API access may not stop immediately after your credits run out," which can leave you with a small negative balance that comes off your next top-up. OpenAI's own words: don't rely on the prepaid balance "as an instantaneous spending cutoff."

What transcription costs.

ModelPer minutePer hour
gpt-4o-mini-transcribe$0.003$0.18
gpt-transcribe$0.0045$0.27
gpt-4o-transcribe$0.006$0.36
whisper-1$0.006$0.36

gpt-transcribe is new on the pricing page and isn't in Humla's model list yet. Humla offers whisper-1, gpt-4o-mini-transcribe, gpt-4o-transcribe and gpt-4o-transcribe-diarize.

What summaries cost. This is the bit Groq and Deepgram can't do. Humla's default summary model on OpenAI is gpt-5.4-mini, listed at $0.75 per million input tokens and $4.50 per million output tokens. My own back-of-envelope: an hour of conversation is somewhere around 9,000 words of transcript, call it 12,000 tokens in, plus a summary out. That works out at a few cents per meeting. Treat that as arithmetic, not a quote; long reasoning or a very long meeting moves it.

And it's one key for both. Humla keeps one OpenAI key in your Keychain and uses it for transcription, summaries and chat over your notes, wherever you've picked OpenAI.

What a month actually costs

Say you have 20 hours of meetings a month, a fairly normal consultant's calendar. Transcription only:

Provider + model20 hours of audio
Groq whisper-large-v3-turbo~$0.80
OpenAI gpt-4o-mini-transcribe~$3.60
Deepgram Nova-3 monolingual~$5.16 (from free credit for a long while)
OpenAI whisper-1~$7.20
On-device Whisper$0

That's per hour of audio sent, which isn't quite the same as hours in the calendar: apps skip silence, and Humla handles your mic and the call audio as separate sources. Use the table for the order of magnitude, which is the useful part anyway. Under a dollar to single digits, a month.

Now the bottom row. That's not a joke row. Humla ships on-device Whisper models that run on Apple Silicon and send nothing anywhere, and for a lot of people that's simply the answer. The cloud providers earn their place when you want a faster result, or a model that's better at a particular language. Humla lets you mix: a per-language override can send, say, English to Deepgram and Norwegian to a local model.

Putting the key into Humla

  1. Settings → Transcription. Pick the provider and model. The key field for that provider appears right below.
  2. Paste the key, press Save, then Test. Humla makes one cheap authenticated request to the provider and shows Connected ✓ or the provider's error.
  3. For summaries, Settings → Summary has its own OpenAI (and Anthropic) key field. Same OpenAI key, same Keychain slot, so if you saved it under Transcription it's already there.

Keys live in the macOS Keychain, not in a settings file. Keys for providers you only use in a per-language override go under Advanced → Other API keys.

The honest version

If you just want transcripts and you're on Apple Silicon, start with no key and see if on-device is good enough for you. It costs nothing, it's private by construction, and it's the reason I built the thing.

If you want cloud, get the Groq key first. No card, four cents an hour, done in two minutes. Get the OpenAI key when you want summaries without running a local model, and turn off that auto-reload toggle on the way through.

And whichever you pick, read what they do with your audio. That's the bit the pricing pages don't put in the table.

Try Humla

Download Humla for macOS - free, MIT licensed, no account. Bring a key from any of the three above, or none. The source is on GitHub if you'd rather read what happens to your key than take my word for it.

Related reading: the best free meeting transcription for Mac, how to transcribe after the meeting instead of during it, and what "private" actually means for meeting notes.