How much does AssemblyAI cost?

AssemblyAI charges per hour of audio you send, not per seat or per month. On pay as you go, Universal-2 lists at 0.15 dollars per hour ¹, and the more accurate Universal-3.5 Pro lists at 0.21 dollars per hour ². There is no minimum commitment, upfront fee or contract, and the company bills monthly for the previous month's usage ³. These are list prices read on its pricing page on 27 October 2026, and they change.

The price per hour depends on the model and the product

Pick the model first, because it sets the rate. The table lists the main products on the pricing page.

For most recorded audio the choice is between the first two rows. The 0.06 dollar gap per hour buys the company's most accurate batch model, but it covers 18 languages, against 99 for Universal-2 ^{7 8}. If your audio is in a language outside the 18, the documentation says it falls back to Universal-2.

Add-ons and channel counts change the bill

The headline rate is not always the final rate. Two things move it most.

• Speaker labels. Diarization is an add-on at 0.02 dollars per hour ⁹. Universal-3.5 Pro with speaker labels therefore comes to 0.23 dollars per hour ¹⁰.
• Stereo files. Multichannel audio is billed per channel, so a one-hour stereo file counts as two hours of transcription ¹¹. A podcast recorded with the host and guest on separate tracks costs twice as much as the same hour mixed to one track. If you do not need separate channels, mix down before you upload.

Other add-ons, such as keyterms prompting for product names and jargon, are listed on the pricing page with prices that depend on the model. Read that page for the exact add-on you plan to use.

The free tier covers a prototype, and it is cheaper than the main rivals

New accounts start on a free tier of up to 185 hours of pre-recorded transcription and up to 333 hours of streaming ¹². The question is AssemblyAI free covers what that includes and when it runs out.

On price alone, AssemblyAI sits at the low end of the hosted APIs. OpenAI's whisper-1 API costs 0.36 dollars per hour of audio at its listed per-minute rate ¹³, so Universal-3.5 Pro is about 42 percent cheaper ¹⁴. Deepgram's Nova-3 lists at 0.258 dollars per hour for pre-recorded audio on pay as you go ¹⁵. Our comparison of Deepgram and AssemblyAI and the ranking of the cheapest speech-to-text APIs go further.

When the cheapest rate is not the cheapest bill

List price is one input. These cases change the answer:

• Accuracy costs time. A model with fewer errors means less manual correction. For transcripts that a person must review, a few cents per hour matters less than the editing time saved.
• Volume. The rates above are pay as you go. Any different terms at high volume are not on the public rate list, so ask the company directly if your usage is large.
• Self-hosting. Open models such as Whisper have no per-hour fee, but you pay for hardware and engineering time. See the guide to running Whisper locally.
• Privacy. If audio cannot leave your device, no cloud price applies. See our question on whether speech-to-text APIs store audio.

What to do before you commit

1. Estimate your monthly audio hours, and count a stereo file as two hours.
2. Multiply by the rate for the model you need, and add 0.02 dollars per hour if you need speaker labels.
3. Run 30 to 60 minutes of your own audio through the free tier with both batch models, and compare errors as well as price. The guide to AssemblyAI describes the test.
4. Check the pricing page again before you launch. Our overview of speech recognition shows other options, including on-device dictation such as Yap, which we make and which has its own pricing.

For the measure behind these accuracy claims, see the glossary entry on word error rate.