TL;DR verdict: Deepgram and AssemblyAI are the two speech-to-text APIs most developers end up choosing between. AssemblyAI is cheaper and simpler for recorded audio — $0.15/hour on Universal-2 and $0.21/hour on Universal-3.5 Pro, with flat per-second billing and a deep audio-intelligence layer. Deepgram costs more per minute ($0.0043/min, about $0.26/hour, for pre-recorded Nova-3) but is built for real-time: low-latency streaming, a voice-agent model, and text-to-speech under one roof. Batch transcription and analysis: AssemblyAI. Live voice products: Deepgram.
If you are building anything that turns speech into text — a meeting tool, call analytics, a voice agent — the shortlist usually ends at Deepgram vs AssemblyAI. Both are API-only and accurate, and both advertise a low headline rate that add-ons push upward. The real differences are pricing shape, streaming strength, and how much intelligence you get on top of the transcript. Research-based comparison, rates checked against current 2026 pricing sources, no affiliate links.
Deepgram vs AssemblyAI at a glance
| Factor | Deepgram | AssemblyAI | Edge |
|---|---|---|---|
| Flagship model | Nova-3 (plus Flux for voice agents) | Universal-3.5 Pro (plus Universal-2 budget tier) | Split |
| Pre-recorded price | $0.0043/min (~$0.26/hr), monolingual Nova-3 | $0.15/hr Universal-2; $0.21/hr Universal-3.5 Pro | AssemblyAI |
| Streaming price | Nova-3 from $0.0048/min promotional (regular $0.0077/min) | $0.45/hr (~$0.0075/min) Universal-3.6 Pro Realtime | Deepgram at promo rates |
| Free starting credit | $200 credit | $50 credit (about 333 hours of Universal-2) | Deepgram on dollars, AssemblyAI on hours |
| Billing style | Per minute, varies by model, mode and add-ons | Per second, flat per-hour rates, no minimums | AssemblyAI for predictability |
| Speaker diarization | Included on pre-recorded; +$0.0020/min streaming | +$0.02/hr add-on (async) | Deepgram pre-recorded, AssemblyAI streaming |
| Audio intelligence | Sentiment, topics, summarization as add-ons | Core strength: sentiment, entities, chapters, summaries, PII redaction, LeMUR LLM layer | AssemblyAI |
| Beyond transcription | Aura-2 text-to-speech ($0.030 per 1,000 characters), Voice Agent API | Text-to-speech API, medical mode, LLM analysis | Deepgram for voice agents |
| Discount route | Growth plan from $4,000/yr commitment; pre-recorded drops to $0.0036/min | Enterprise volume discounts at high monthly hours | Split |
How does the pricing actually work?
AssemblyAI prices like a utility: one flat rate per hour of audio, billed per second, no minimums and no contract. Universal-2 at $0.15/hour is the cheapest serious transcription in this matchup, and Universal-3.5 Pro at $0.21/hour is the accuracy tier. Add-ons are itemized per hour — diarization adds $0.02/hour, PII redaction $0.03/hour, and extras like Medical Mode add $0.15/hour — so a fully loaded hour can land near $0.40, but you can predict it on a napkin. Our AssemblyAI Review 2026 walks through the full add-on math.
Deepgram prices per minute, and the rate depends on the model, whether audio is pre-recorded or streaming, and whether it is monolingual or multilingual. Nova-3 pre-recorded monolingual is $0.0043/minute (about $0.26/hour); streaming runs higher, with promotional rates below the regular $0.0077/minute list price. Add-ons such as streaming diarization layer on top. It is still cheap in absolute terms, but forecasting a Deepgram bill takes a spreadsheet where AssemblyAI takes a sentence. Our Deepgram Review 2026 breaks down its tiers and add-ons in detail.
Which is better for real-time and voice agents?
Deepgram. Streaming is its home turf: Nova-3 streams with low latency, and the Flux model is built specifically for voice-agent turn-taking — knowing when a caller has finished speaking is half the battle in a phone agent. Deepgram also sells the full loop in one place, pairing transcription with Aura-2 text-to-speech and a Voice Agent API, which saves stitching vendors together.
AssemblyAI does stream — Universal-3.6 Pro Realtime at $0.45/hour — and it is a credible option for live captions and agent assist. But its product gravity is asynchronous: upload audio, get back a rich, analyzed transcript. If your product talks back in real time, Deepgram is the more natural foundation.
Which gives you more than a transcript?
AssemblyAI, clearly. Its intelligence layer is the product: speaker labels, sentiment analysis, entity detection, auto chapters, summaries, content moderation, and PII redaction, plus LeMUR for asking an LLM questions about the audio. For call-center analytics, podcast pipelines, and media archives, that consolidation replaces several services at once.
Deepgram offers intelligence features too — sentiment, topic detection, summarization — but they read as add-ons to a transcription engine rather than the main event. If you only need fast, accurate words, that focus is a virtue. If you need to understand the conversation, AssemblyAI starts ahead. For context on how these APIs differ from end-user transcription apps, see our Otter.ai vs Trint comparison.
Which should you pick?
Pick AssemblyAI if you transcribe recorded audio at volume — meetings, podcasts, calls, archives — and want the lowest predictable price plus built-in analysis. Start on Universal-2; step up to Universal-3.5 Pro where accuracy pays for itself.
Pick Deepgram if latency is the product: voice agents, live captioning, or real-time call features, especially if you want speech-to-text and text-to-speech from one vendor. Use the $200 starting credit to benchmark it on your own audio before committing.
Either way, test both on your real recordings before a big commitment. Accent mix, background noise, and industry vocabulary move accuracy more than any published benchmark, and both vendors give you enough free credit to find out.
Frequently asked questions
Is Deepgram or AssemblyAI cheaper?
AssemblyAI for recorded audio: $0.15/hour on Universal-2 and $0.21/hour on Universal-3.5 Pro, versus about $0.26/hour for Deepgram’s pre-recorded Nova-3. Streaming comparisons depend on Deepgram’s current promotional rates.
Which speech-to-text API is better for voice agents?
Deepgram. Its Flux model is designed for voice-agent turn detection, streaming latency is its core strength, and its Voice Agent API plus Aura-2 text-to-speech cover the full conversation loop.
Do Deepgram and AssemblyAI offer free tiers?
Both offer starting credit rather than permanent free plans: Deepgram gives $200 in credit, and AssemblyAI gives $50 — roughly 333 hours of Universal-2 transcription.
What do add-ons do to the real price?
They can add 30–100% on either platform. AssemblyAI itemizes them per hour (diarization +$0.02/hr, PII redaction +$0.03/hr); Deepgram itemizes per minute (streaming diarization +$0.0020/min). Price your full configuration, not the headline rate.
Prices checked October 2026 against current published rates; promotional streaming rates change, so confirm on each vendor’s pricing page before budgeting. Research-based comparison — no hands-on testing is claimed.