GuidesVoice providers
Cartesia AI: Sonic and Ink pricing, latency and your own AWS account
Call 1800 570 693 to talk to the demo agent. Plans start at A$299 a month, GST included.
- Voice, on the Scale plan
- About USD 0.034 per spoken minute
- Line voice agents
- USD 0.06 a minute
- In your own AWS account
- Sonic 3 on SageMaker, in Sydney
- First audio, measured from Sydney
- 448 ms median, 604 ms p90
Prices read on cartesia.ai on 6 October 2026. SageMaker region read from the AWS API the same day. Latency measured by Vertical AI on live phone calls from Sydney over the 14 days to 4 October 2026.
Cartesia models
Cartesia builds its models on state space models, a different architecture from the transformers most voice models use. Sonic 3.6, generally available since 27 August 2026, is the current voice model.
Ink-2 transcribes the caller and reports turns as it goes: turn.start, turn.update, turn.eager_end, turn.resume and turn.end. An agent built on it needs no separate voice activity detector to decide when the caller has stopped.
| Model | Use | Languages | Cartesia's latency claim |
|---|---|---|---|
| Sonic 3.6 | Text to speech | 44, across 61 locales | Under 90 ms |
| Sonic 3.5 | Text to speech | 42 | Not published separately |
| Sonic 3 | Text to speech, on SageMaker | 42 | Under 100 ms |
| Ink-2 | Speech to text with turn detection | English, French, Hindi, Japanese, Spanish | 0.1 s to a final transcript |
| Ink Whisper | Speech to text, lowest price | About 100 | Not published |
Model retirements on 20 October 2026
Cartesia's docs say sonic-turbo, sonic-2 and the sonic-3-2025-10-27 snapshot stop working after 20 October 2026. An agent pinned to any of them needs moving to Sonic 3.5 or 3.6 before that date.
Cartesia pricing
Cartesia sells monthly plans in US dollars, each with a pool of credits. Text to speech uses one credit per character, and Ink uses three credits per second of audio.
Line voice agents cost USD 0.06 a minute. A phone number bought from Cartesia adds USD 0.014 a minute, and you can bring your own Twilio number or SIP trunk instead.
| Plan | Price a month | Credits | Concurrent voice streams | Concurrent transcriptions |
|---|---|---|---|---|
| Free | USD 0 | 20,000 | 2 | 8 |
| Pro | USD 5 | 100,000 | 3 | 12 |
| Startup | USD 49 | 1,250,000 | 5 | 20 |
| Scale | USD 299 | 8,000,000 | 15 | 60 |
| Enterprise | Quoted | Quoted | Quoted | Quoted |
Cost per spoken minute
On the Scale plan a credit costs about USD 0.0000374. At 900 characters a spoken minute, Sonic costs about USD 0.034 per minute of agent speech.
Ink uses 180 credits per minute of caller audio. Cartesia prices that at USD 0.39 an hour on Scale, about USD 0.0065 a minute. Voice streams are the tighter limit: Scale allows 15 at once, so 15 callers can hear the agent at the same moment.
Where Cartesia processes Australian data
The public endpoint, api.cartesia.ai, sends each request to the nearest Cartesia region. Cartesia publishes no Australian address for it, so we treat it as overseas.
Cartesia lists Australia among its dedicated regional deployments, beside the United States, the European Union, the United Kingdom and India. They are Enterprise only, and Cartesia gives out the hostname on request through support@cartesia.ai.
Zero data retention is also Enterprise only. It covers text to speech and speech to text inputs and outputs, but not voice cloning, and Cartesia still keeps request IDs and usage totals.
Cartesia latency measured from Sydney
We measure time to first audio on every turn of every live call, from sending text to the voice provider to receiving the first audio back. Our call servers run in Sydney.
Over the 14 days to 4 October 2026, Cartesia took 448 ms at the median and 604 ms at the 90th percentile. ElevenLabs took 499 ms and 842 ms on the same servers, a gap of 238 ms at the 90th percentile.
| Voice provider | Median first audio | 90th percentile |
|---|---|---|
| Cartesia | 448 ms | 604 ms |
| ElevenLabs | 499 ms | 842 ms |
Three ways to buy Cartesia through AWS
Cartesia can be bought through AWS three ways. Two of them move the processing into your own AWS account.
| Route | Where audio is processed | What it covers | How it is billed |
|---|---|---|---|
| AWS Marketplace subscription | Cartesia's cloud | Sonic, Ink and Line | USD 299 a month on Scale, through your AWS bill |
| SageMaker JumpStart | Your AWS account, Sydney available | Sonic 3 voice only | An ml.g6e.xlarge GPU instance by the hour, plus any Cartesia software fee |
| Enterprise self-hosted | Your VPC or data centre, air-gapped available | Sonic and Ink-2 | Quoted by Cartesia |
Sonic on SageMaker
AWS added Sonic 3 to SageMaker JumpStart on 4 February 2026. On 6 October 2026, the JumpStart catalogue listed cartesia-sonic-3-sagemaker version 1.0.4 in ap-southeast-2, Sydney.
Cartesia's own deployment repository also covers Sonic 3.5 and 3.6, and recommends an ml.g6e.xlarge instance. The SageMaker endpoint streams audio in a different format from Cartesia's public WebSocket, so a voice agent needs a client written for it.
Ink is not on JumpStart. A deployment that must keep the caller's audio inside your account needs Cartesia's Enterprise self-hosted product for transcription, or a different speech to text provider such as Deepgram's self-hosted Nova-3.
Cartesia compliance
Cartesia lists SOC 2 Type II, HIPAA, GDPR and PCI DSS as a service provider on its trust centre. It signs business associate agreements and data processing agreements on Enterprise. We found no ISO 27001 or IRAP certification.
Where Cartesia fits an Australian deployment
Our voice agents use Cartesia Ink-2 for speech recognition by default, and offer Sonic voices where an agent needs the fastest reply. Cartesia publishes no Australian address on its public API, so both run overseas unless you buy the Enterprise Australian deployment.
For a buyer who needs everything in Australia without an Enterprise contract, Deepgram's Sydney endpoint is the alternative. For a buyer who wants the voice inside its own AWS account, Sonic on SageMaker in Sydney is the shortest route.
Questions
Cartesia plans cost USD 0 on Free, USD 5 on Pro, USD 49 on Startup and USD 299 on Scale each month, and Enterprise is quoted. Text to speech uses one credit per character, about USD 0.034 per spoken minute on Scale. Line agents cost USD 0.06 a minute. Prices read on cartesia.ai on 6 October 2026.
Anyone in Australia can use Cartesia's public API, which routes to the nearest Cartesia region and publishes no Australian address. Cartesia sells a dedicated Australian deployment on Enterprise, and Sonic 3 runs on SageMaker in Sydney inside your own AWS account.
Cartesia sells a Marketplace subscription at USD 299 a month for the Scale plan, billed through your AWS account, with audio still processed by Cartesia. Sonic 3 is also a SageMaker JumpStart model that runs on a GPU instance in your own account.
Sonic is Cartesia's text to speech model. Sonic 3.6, generally available since 27 August 2026, speaks 44 languages, and Cartesia advertises under 90 ms to first audio. From Sydney we measured 448 ms at the median, including the network.
Ink is Cartesia's speech to text. Ink-2 streams transcripts with turn events built in, in English, French, Hindi, Japanese and Spanish. Ink Whisper covers about 100 languages at a lower price.
On our live calls from Sydney over the 14 days to 4 October 2026, Cartesia returned first audio in 448 ms at the median and 604 ms at the 90th percentile. ElevenLabs took 499 ms and 842 ms on the same servers.
Cartesia sells Sonic and Ink-2 for on-premises, air-gapped and private cloud deployment on AWS, Google Cloud or Azure, on an Enterprise contract. Cartesia quotes these deployments directly.
Sources
- Cartesia pricing
- Cartesia Sonic 3.6 announcement
- Cartesia Ink
- Cartesia text to speech models
- Cartesia older models and retirements
- Cartesia speech to text models
- Cartesia regional endpoints
- Cartesia zero data retention
- Cartesia trust centre
- AWS: Cartesia Sonic 3 on SageMaker JumpStart
- Cartesia on AWS Marketplace
- cartesia-ai/cartesia-aws on GitHub
Related pages
Get started
Tell us what your callers ring about, and we build and test the agent for you.