Head-to-head
Deepgram vs AssemblyAI: which one in 2026?
Deepgram and AssemblyAI are both developer-facing speech-to-text APIs. Deepgram leans on live conversation models, regional endpoints and self-hosting, while AssemblyAI, which sells only a hosted API, publishes per-hour rates for each model and spells out its retention and deletion windows in a dedicated document.
Verdict
Deepgram sits ahead because it covers more deployment cases: a global and regional cloud, self-hosting, and models aimed at live conversation. Its weak point is the default data setting, since audio is retained unless you opt out. AssemblyAI is the easier one to budget and to explain to a compliance reviewer. Pick Deepgram for agents and control over where audio runs, AssemblyAI for batch work with clear numbers.
Deepgram and AssemblyAI side by side
| Deepgram | AssemblyAI | |
|---|---|---|
| Price entry | $0.015 per 1,000 characters (Aura-1, pay as you go) | $0.15 per hour |
| How you start | Pay as you go begins with a $200 credit that does not expire, requires no card, and carries no minimums. | A free tier gives $50 in credits without a card, limited to 5 new streams per minute until you upgrade. |
| Published rates | Billing units differ by product (per second, per 1,000 characters or per connection minute), and the pricing page shows the units rather than one hourly figure. | Rates are listed per hour: Universal-3.5 Pro pre-recorded at $0.21, Universal-2 at $0.15, and Universal-Streaming at $0.15, as of November 2026. |
| Real-time speech | Flux STT targets live conversation with turn detection, interruption handling and low latency, which suits voice agents. | Universal-Streaming and Universal-3.6 Pro Realtime are billed by connection hour, with the Realtime model at $0.45 per hour base. |
| Domain vocabulary | Medical, pharma, legal and finance models exist, plus keyterm prompting and custom models trained on your own datasets. | Keyterms, custom spelling and prompting steer output, and a medical mode handles clinical terminology in transcripts. |
| Concurrency | The Growth plan includes 225 streaming connections; Enterprise sets concurrency to match your expected peak traffic. | Pay-as-you-go allows 100 new streams per minute; higher ceilings come only through an enterprise contract. |
| Default data use | Model Improvement Program is on by default; adding mip_opt_out=true stops storage of audio and transcripts after the response. | Files are excluded from training once you opt out. Zero retention covers streaming and sync requests for opted-out customers. |
| Where audio is processed | EU, Australia and India endpoints exist, and a full in-region guarantee needs the regional endpoint plus the opt-out. | Files processed on European servers are excluded from training. The pages checked do not list other regions. |
| Deployment options | Runs on Deepgram cloud, a regional endpoint, or inside your own environment through self-hosting on Enterprise. | The pages checked describe a hosted API only; a self-hosted option was not found on them. |
Where each one wins
Deepgram wins on
- Self-hosted deployment and an EU endpoint for teams that cannot send audio to a shared global API
- Flux and custom models trained on your data for voice agents and unusual vocabularies
- Credit that never expires, so a prototype can sit idle without losing its balance
AssemblyAI wins on
- Per-model hourly rates are printed on the page, which makes cost forecasting a spreadsheet exercise
- Retention is documented in numbers, such as 24 to 48 hour audio deletion and 30 day transcript deletion
- LLM Gateway and text-to-speech sit under one account for pipelines that add summaries or voice output
Which to pick
Pick Deepgram if
You are building a live voice agent, need self-hosting or an EU endpoint, or want models tuned to medical, legal or finance language.
Pick AssemblyAI if
You transcribe recorded audio at volume, want published hourly rates, and need written retention windows before a security review.
Each tool in context: Deepgram profile, AssemblyAI profile, and the full Best AI Speech to Text APIs in 2026.
Frequently asked questions
Does Deepgram or AssemblyAI train on my audio by default?
Deepgram's Model Improvement Program is the default, so audio and transcripts are retained for training unless each request sets mip_opt_out=true. AssemblyAI says only certain files are used for training and excludes files from it when you opt out, including anything processed on European servers.
Which of the two has clearer pricing for pre-recorded audio?
AssemblyAI. Its pricing page lists Universal-3.5 Pro pre-recorded at $0.21 per hour and Universal-2 at $0.15 per hour as of November 2026. Deepgram bills in different units by product, so you need to convert before comparing, and Growth starts at $4,000 per year.
Can I run either one inside my own infrastructure?
Deepgram offers self-hosting, listed on its Enterprise plan alongside custom model training and a HIPAA BAA. AssemblyAI's opened pages describe a hosted API with enterprise contracts for volume, and no self-hosted option was found, so ask sales if that is a requirement.
How much free usage do Deepgram and AssemblyAI give new accounts?
Deepgram gives a $200 credit that does not expire and needs no card. AssemblyAI gives $50 in credits, also without a card, but caps the free tier at 5 new streaming sessions per minute until you move to pay-as-you-go, which allows 100.