Summary: Deepgram, AssemblyAI and Gladia all let you transcribe audio through an API with free starting credit and no subscription. Deepgram gives the largest free credit ($200), AssemblyAI has the lowest dollar list prices for both pre-recorded and streaming audio, and Gladia's Solaria-3 model is a similarly low-cost option, billed in euros, for English and four other European languages. Every price and limit below was checked on each company's official pricing page and docs on October 31, 2026.
Who this comparison is for
If you are building a meeting-notes tool, a podcast subtitle pipeline, a call-center dashboard or a voice agent, you will eventually pick a speech-to-text (STT) API. The three providers here are popular because you can sign up, get an API key and transcribe your first file in a few minutes without talking to sales. Their pricing pages, however, use different units (per minute, per hour, euros, dollars), and the free credits behave differently. This post puts them side by side in the same units and adds the setup gotchas the docs warn about, so you can make a first choice without opening a dozen tabs.
One honest limit: this is a comparison of published prices, limits and documented behavior. I did not run my own accuracy benchmark, so treat any "most accurate" claim on vendor pages as marketing until you test with your own audio.
The free credit, explained for each
Deepgram: $200 that does not expire
Deepgram's Pay as you go plan starts with a $200 credit, no credit card required, and the pricing FAQ says the credit does not expire. Deepgram estimates it covers roughly 43,000 minutes (over 700 hours) of Nova-3 transcription. After that you keep paying per second of audio with no minimums. A Growth plan exists for heavier users: it starts at $4,000 a year in prepaid credit and lowers rates by up to 20%.
AssemblyAI: $50, with tighter concurrency until you pay
AssemblyAI's pricing reference lists $50 in free credits at signup, also with no credit card. The catch is concurrency. The rate-limits doc says a free account can process 5 transcriptions in parallel (paid accounts start at 200+), and free accounts can open only 5 new streaming connections per minute versus 100 on pay-as-you-go. Extra jobs are queued, not rejected, so a bulk upload on a free account simply takes longer.
Gladia: €50, one-time
Gladia gives new accounts €50 in credits. The getting-started doc says they do not renew once used, and the pricing FAQ adds that they are valid for one year. Gladia estimates that is roughly 312 hours of async transcription on Solaria-3, 185 hours on Solaria-1, or 73 hours of real-time. Free accounts are limited to 3 concurrent pre-recorded jobs and 1 live session.
Prices side by side
All three are usage-based, so there is no monthly-versus-annual toggle to watch for on the pay-as-you-go tier. The only annual commitment is Deepgram's Growth plan, which I left out of the table. I converted Deepgram's per-minute prices to per-hour so the rows line up (multiply by 60). Gladia prices in euros, so compare it with your own exchange rate.
| Deepgram | AssemblyAI | Gladia | |
|---|---|---|---|
| Free start | $200 credit, no card, no expiry | $50 credit, no card | €50 one-time credit, valid 1 year |
| Pre-recorded (list price) | Nova-3 monolingual $0.0043/min (about $0.26/hr); multilingual $0.0052/min | Universal-3.5 Pro $0.21/hr; Universal-2 $0.15/hr | Solaria-3 €0.16/hr; Solaria-1 €0.27/hr (first 100 hrs/month) |
| Streaming (list price) | Nova-3 or Flux English $0.0077/min (about $0.46/hr) | Universal-Streaming $0.15/hr; Universal-3.6 Pro Realtime $0.45/hr | €0.68/hr |
| Speaker labels | Included for pre-recorded; $0.0020/min for streaming | +$0.02/hr (async standard); +$0.12/hr streaming | Included, async only |
| Free-tier concurrency | Pay as you go: 150 streaming connections | 5 parallel jobs; 5 new streams per minute | 3 async jobs; 1 live session |
| Volume discounts | Growth plan (prepaid yearly, up to 20% off) | Custom pricing via sales | Automatic by monthly hours, plus top-up bonuses from €1,000 |
What 100 hours a month would cost
To make the table concrete, here is the arithmetic for 100 hours of audio in one month at the list prices above (before free credit, add-ons excluded unless noted):
- Pre-recorded, English: Deepgram Nova-3 about $25.80 (6,000 minutes × $0.0043), speaker labels included. AssemblyAI Universal-3.5 Pro $21, or $23 with standard speaker labels; Universal-2 $15. Gladia Solaria-3 €16, or Solaria-1 €27.
- Streaming: Deepgram Nova-3 monolingual about $46.20. AssemblyAI Universal-Streaming $15, or $45 on Universal-3.6 Pro Realtime. Gladia €68.
Two billing rules change these numbers in real projects. First, Deepgram and AssemblyAI both bill multichannel audio per channel, so a one-hour stereo call is charged as two hours unless you downmix it. Second, AssemblyAI bills streaming by how long the WebSocket session stays open, not by how much audio you send. Its pricing reference gives the example of a connection left open for 60 minutes with 30 minutes of audio: you pay for 60. Deepgram, by contrast, bills per second of processed audio with no rounding to blocks.
Your first transcription on each
These are the shortest paths from the official quickstarts. Each needs an API key from the provider's dashboard first.
Deepgram (cURL, no SDK)
- Sign up and create an API key in the Deepgram console.
- Send a POST request to
https://api.deepgram.com/v1/listen?model=nova-3&smart_format=truewith the headerAuthorization: Token YOUR_DEEPGRAM_API_KEY. - For a hosted file, send JSON like
{"url":"https://dpgr.am/spacewalk.wav"}withContent-Type: application/json. For a local file, send the raw bytes with--data-binary @youraudio.wavand a matching audio content type.
Keep model=nova-3 in the URL. The quickstart notes that if you remove the parameter, Deepgram falls back to its older base model.
AssemblyAI (Python SDK)
- Copy your key from the dashboard and run
export ASSEMBLYAI_API_KEY=<your-key>. - Install the SDK with
pip install assemblyai(Python 3.8+), ornpm install assemblyaifor Node.js 18+. - Create a
Transcriberand calltranscribe()with a URL or a local path. The SDK uploads, submits and polls for you, then you printtranscript.text.
Set the model explicitly with speech_models (plural). AssemblyAI's own docs warn that defaults can differ between free and paid accounts, and that the singular speech_model is deprecated and returns a 400 error with current model names.
Gladia (Python SDK)
- Sign up at app.gladia.io and open the API keys page. The docs say a default key is already created for you.
- Install the SDK with
pip install gladiaio-sdk(ornpm install @gladiaio/sdk). - Create a
GladiaClientwith your key, call.prerecorded(), thentranscribe()with a URL or local file.
If you don't pass model, Gladia uses Solaria-1. To get the cheaper Solaria-3 rate you must set model: "solaria-3" and pass exactly one language in language_config.languages. The quickstart lists English, French, German, Spanish and Italian, says it is async only, and says it does not support code switching.
Which one should you pick?
- You want the longest free runway for prototyping: Deepgram. $200 with no expiry is four times the other two credits, and pre-recorded speaker labels are included.
- You run cheap, high-volume real-time captions: AssemblyAI Universal-Streaming at $0.15/hr has the lowest streaming list price here, as long as you close sessions promptly.
- You transcribe English or major European-language recordings in batches: Gladia Solaria-3 at €0.16/hr, with automatic volume discounts that drop to €0.12/hr above 10,000 hours a month.
- You need many languages in one file: look at Gladia Solaria-1 (100+ languages, code switching) or Deepgram's multilingual Nova-3, and test both on your real audio.
Editor's note
Good fit: developers who can write a small script and want to pay only for the audio they process. All three work without a subscription.
Not ideal: people who just want a transcript of one meeting without code. A ready-made notetaker app is easier. Also, if your audio contains sensitive data, read each provider's data-use terms first. Gladia's pricing page, for example, says self-serve accounts are opted in to model training by default, with a free opt-out in your profile.
Common errors and fixes (from the official docs):
- Deepgram 400 "corrupt or unsupported data": usually you sent a JSON body with a URL but forgot
Content-Type: application/json. If you sent a file, check it withffprobe. - Deepgram 402 ASR_PAYMENT_REQUIRED: the project has run out of credit. Add funds or a card.
- AssemblyAI job "succeeds" but the transcript shows
status: error: the submit call can return 200 and the job can still fail later. Always check theerrorfield when you poll. - AssemblyAI 429 while polling: you hit the limit of 60 GET requests per minute for one transcript. Wait for the
Retry-Afterheader, or use webhooks instead of tight polling loops. - Gladia 429: concurrency is used up (3 jobs on free). Back off and retry. Do not resubmit a job that already returned 200. Gladia says that creates a new billable job and can make the backlog worse.
FAQ
Do I need a credit card to try these APIs?
No. Deepgram and AssemblyAI both say no card is required for the free credit, and Gladia's free tier is a one-time €50 credit grant. You only add a payment method when you top up or want higher limits.
Which is cheapest for pre-recorded English audio?
By list price, AssemblyAI Universal-2 ($0.15/hr) and Gladia Solaria-3 (€0.16/hr) are the lowest, followed by AssemblyAI Universal-3.5 Pro ($0.21/hr) and Deepgram Nova-3 (about $0.26/hr). Add-ons and accuracy on your audio can change which one is cheapest overall, so run a small test batch first.
Why is my streaming bill higher than my audio length?
On AssemblyAI, streaming is billed for the time the WebSocket stays open, including silence and idle time. Close the connection as soon as a call ends. On Deepgram and AssemblyAI, stereo or multichannel audio is also billed per channel.
Related on this blog: ElevenLabs Pricing: Free vs Starter vs Creator Credits · Tavily vs Exa vs Brave Search API: Free Tiers and Real Costs · Wispr Flow vs Aqua Voice vs Willow: AI Dictation Compared
Start with whichever free credit matches your audio, transcribe the same ten real files on all three, and let your own results decide.
Image credits: the author's own screenshots of the public Deepgram, AssemblyAI and Gladia pricing pages, plus the API keys screenshot from Gladia's official documentation. See the official screenshots at the end of this post.
Last verified: 2026-10-31 · By zeromy
Sources:
Deepgram pricing · Deepgram pre-recorded quickstart · Deepgram errors
AssemblyAI pricing · AssemblyAI pricing reference · AssemblyAI quickstart · AssemblyAI rate limits · AssemblyAI common errors
Gladia pricing · Gladia getting started · Gladia pre-recorded quickstart · Gladia concurrency limits · Gladia retry policy
댓글
댓글 쓰기