by AssemblyaiReleased 22 September 2026
Universal-3.5 Pro turns an audio file of up to 20 minutes into a finished transcript with a timestamp on every word, handed back in one reply. Send WAV, MP3, M4A, AAC, FLAC, WebM or Ogg, and add a short instruction, key terms or background about the conversation to steer it toward your names and jargon.
You pay for what each request uses. No subscription, no minimum.
One minute of audio, the ordinary use
Every rate
| Meter | ↑ you send | ↓ comes back |
|---|---|---|
| Second of audio | 0,028 DA |
per second of audio sent
One company runs this model. OpenDunes sends every request to it.
| Provider | Per second of audio | Latency | Uptime |
|---|---|---|---|
| AssemblyAIUnited States | 0,028 DA | … | … |
You pay the price of the provider that served your request. Prices read 15:19 · speed over the last 30 minutes · uptime over 3 days.
Swap the base URL and the key. Your OpenAI code runs unchanged.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://opendunes.com/v1",
api_key=os.environ["OPENDUNES_API_KEY"],
)
with open("meeting.mp3", "rb") as f:
resp = client.audio.transcriptions.create(model="assemblyai/universal-3-5-pro", file=f)
print(resp.text)Runs this request once on your balance.
About 1,66 DA per minute of audio. ↑ 0,028 DA per second of audio.
No. You pay per use from your balance, with no subscription. Try it in Chat at the same price.
Universal-3.5 Pro turns an audio file of up to 20 minutes into a finished transcript with a timestamp on every word, handed back in one reply. Send WAV, MP3, M4A, AAC, FLAC, WebM or Ogg, and add a short instruction, key terms or background about the conversation to steer it toward your names and jargon.
Universal-3.6 Pro costs the same for the same minute of audio.
Deposit dinars with your Edahabia or CIB card. Every model, Universal-3.5 Pro included, is paid from that balance. No foreign card, no conversion.
AssemblyAI states 18 languages for Universal-3.5 Pro, and the list it publishes for the streaming edition includes French and Arabic. Darja is not named.
Speaker labels: not stated. Word timings: yes.
Up to 25 MB as an upload. Longer audio goes in the request body.
Assemblyai. Released 22 September 2026.
Universal-3.6 Pro turns an audio file of up to 20 minutes into a transcript with a timestamp on every word, handed back in a single reply. AssemblyAI says it makes 30% fewer mistakes on brief back-and-forth speech and 12% fewer across languages on average, adds 14 languages, and writes out phone numbers, street addresses, emails and codes more accurately. You can add key terms or background to steer its vocabulary.
This model takes in written text and gives back spoken audio, so a buyer who has a script or a set of lines to be read aloud is its user. It is built for voice assistants that answer callers, for narrated recordings, and for speech in several languages, and it draws on a shared catalogue of thousands of ready-made voices to pick from. It cannot make a voice from a sample of someone's recording, so a job that must sound like one particular person falls outside what it does.
From what its maker and two benchmark sites publish, and where it stands on OpenDunes. Read 7 Nov 2026.
Word timestamps
Gives the time each word was spoken. Read in the maker's description.
Transcribe a short customer voicemail
Get product names right with a key-term list
Transcribe a voice command for your app