AssemblyAI prices speech to text by the hour of audio: $0.15 to $0.21 per hour for pre-recorded files and $0.15 to $0.45 per hour for streaming. Add-ons such as speaker diarization, PII redaction and entity detection each add a few cents per hour on top. New accounts get $50 in free credits that do not expire.
Those figures were read from AssemblyAI’s pricing page and its billing documentation on November 10, 2026. AssemblyAI also sells a text-to-speech API and an LLM Gateway, which are priced separately and covered below.
We build Cartesia, whose Ink models compete with AssemblyAI’s streaming speech to text, so we have an interest here. Every AssemblyAI number below comes from AssemblyAI’s own pages. Our prices are at the end so you can check both.
AssemblyAI plans
| Plan | Price | What changes |
|---|---|---|
| Free | $50 credit, no card | 5 new streaming sessions per minute; credits cover STT, not LLM Gateway |
| Pay as you go | Per-hour usage | 100 new streaming sessions per minute; no contract or minimum |
| Enterprise | Custom | Higher concurrency, volume pricing, HIPAA BAA at no premium |
AssemblyAI says SOC 2 Type 2, ISO 27001 and GDPR compliance come with no premium. With auto-pay off, API access pauses when the balance reaches $0, so turn auto-pay on for anything in production.
Speech-to-text pricing
| Model | Type | Per hour | Per minute |
|---|---|---|---|
| Universal-3.5 Pro | Pre-recorded | $0.21 | $0.0035 |
| Universal-2 | Pre-recorded | $0.15 | $0.0025 |
| Universal-3.6 Pro Realtime | Streaming | $0.45 | $0.0075 |
| Universal-Streaming English | Streaming | $0.15 | $0.0025 |
| Universal-Streaming Multilingual | Streaming | $0.15 | $0.0025 |
The per-minute column is the hourly rate divided by 60. SLAM-1 is listed as deprecated with no price.
For live voice agents, AssemblyAI points to Universal-3.6 Pro Realtime. Keyterm prompting is included on that model; on Universal-Streaming English it adds $0.04 per hour.
How AssemblyAI bills audio
Pre-recorded audio is prorated to the second. A 90-second file bills as 90 seconds. Failed requests are not billed, and if you transcribe only part of a file with audio_start_from or audio_end_at, only that range bills.
Streaming bills on connection time. AssemblyAI bills for as long as the WebSocket stays open, not for the audio you send. Its FAQ tells you to close connections at the end of each call, because idle time between calls counts. A session left open auto-closes after 3 hours and bills all 3.
That second rule matters more for voice agents than the rate does. An agent that keeps one socket open through hold music, or reuses a connection across calls, pays for the silence.
Add-on pricing
Add-ons bill per hour on top of the model rate. On pre-recorded audio:
| Add-on | Per hour |
|---|---|
| Speaker identification | +$0.02 |
| Speaker diarization | +$0.02 |
| Speaker diarization (experimental) | +$0.065 |
| Translation | +$0.06 |
| Custom formatting | +$0.03 |
| Entity detection | +$0.08 |
| Sentiment analysis | +$0.02 |
| Key phrases | +$0.01 |
| Topic detection (IAB) | +$0.15 |
| Keyterm prompting | +$0.05 |
| Medical mode | +$0.15 |
On streaming, diarization is +$0.12 per hour, medical mode +$0.15, and Voice Focus +$0.10 on Universal-3.6 Pro Realtime. General prompting, in beta, is +$0.05 on Universal-3.5 Pro and Universal-3.6 Pro Realtime. Auto chapters and summarization are listed as deprecated.
Guardrails
| Guardrail | Per hour |
|---|---|
| Profanity filtering | +$0.01 |
| PII audio redaction | +$0.05 |
| PII text redaction | +$0.08 |
| Content moderation | +$0.15 |
Stacking is where an AssemblyAI bill grows. Universal-2 with diarization, entity detection, sentiment and PII text redaction is $0.15 + $0.02 + $0.08 + $0.02 + $0.08 = $0.35 per hour, more than twice the base rate.
Text to speech and LLM Gateway
AssemblyAI’s text-to-speech API costs $15 per million input characters, pay as you go. Cloned voices and EU processing cost the same.
The LLM Gateway resells models from OpenAI, Anthropic, Google and Alibaba per million tokens, for example $1.25 input and $10 output for GPT-5, and $3 input and $15 output for Claude 4.6 Sonnet. Those prices apply to global routing; in-region US or EU routing is 10% higher. The $50 free credit does not cover the gateway.
What a real workload costs
Two examples at the rates on AssemblyAI’s page, November 10, 2026.
Transcribing 100,000 minutes of recorded calls a month. That is 100,000 ÷ 60 = 1,666.7 hours. On Universal-3.5 Pro at $0.21, the bill is $350. Add diarization ($0.02) and PII text redaction ($0.08) and the rate becomes $0.31 per hour, or $516.67.
The ear of a voice agent handling 10,000 call minutes a month. That is 166.7 session hours, if every socket closes when its call ends. On Universal-3.6 Pro Realtime at $0.45, the bill is $75. With streaming diarization at $0.12, it is $0.57 × 166.7 = $95. The language model and the voice bill on top, from AssemblyAI or anyone else, and so does the phone line. If sessions stay open 20% longer than the calls, the speech-to-text line grows by 20% too.
Where Cartesia fits
If you are pricing AssemblyAI for a live voice agent, compare against Universal-3.6 Pro Realtime, the model AssemblyAI recommends for that job.
- Price: Ink-2 costs 3 credits per second, or 10,800 credits per hour. At the Scale plan’s $299 for 8 million credits, that is 10,800 × $299 ÷ 8,000,000 = $0.40 per hour, against $0.45 per session hour for Universal-3.6 Pro Realtime.
- Accuracy: on the independent Artificial Analysis streaming benchmark, Ink-2 has a 3.4% word error rate (October 2026). Every misheard account number or name is a turn the agent has to repair, which costs more than the transcription did.
- Turn detection: Ink-2 predicts the end of each turn itself, so you do not run a separate turn-taking model.
The whole agent can also run on Cartesia. Managed Agents cost $0.03 per minute with speech to text, the voice and orchestration in that rate, so the 10,000-minute agent above comes to $300 before the language model. LLM usage is billed separately, and a Cartesia-provided US number adds $0.014 per minute.
If you already call AssemblyAI’s API, our AssemblyAI migration guide maps its submit-and-poll flow and parameters to Cartesia’s, including the features that have no direct equivalent. The speech-to-text model guide covers the other streaming models worth testing.
Cartesia pricing, for comparison
Cartesia’s plans bill speech to text and text to speech from one pool of credits: 1 credit per character of text to speech, and 3 credits per second of Ink-2 audio.
| Plan | Monthly price | Credits per month | Prepaid agent usage | Overage per 1M credits |
|---|---|---|---|---|
| Free | $0 | 20,000 | $1 | Not available |
| Pro | $5 | 100,000 | $5 | $65 |
| Startup | $49 | 1,250,000 | $49 | $45 |
| Scale | $299 | 8,000,000 | $299 | $38 |
| Enterprise | Custom | Custom | Volume pricing | Custom |
Managed Agents cost $0.03 per minute, drawn from the prepaid agent balance first. Unused credits roll over up to twice the monthly amount. Cartesia also deploys in your VPC, on-prem, or air-gapped.
Try Ink-2 in the playground, or send the same call audio through both APIs and compare the transcripts before you commit to either bill.