Overview
Cartesia builds real-time voice models and the agents on top of them, all running on its own state space model research. The product splits into three layers: Sonic for speech synthesis, Ink for streaming transcription, and Managed Agents for voice agents on phone calls. All three come through one API rather than separate vendors.
Managed Agents is the layer in focus. It bundles phone numbers, concurrency, call duration and the speech stack into one platform, so an enterprise can put an agent on the phone without building telephony itself. Calls are billed at 0.03 dollars per minute, plus 0.014 dollars per minute when using a Cartesia-provided number, with any LLM called during the call billed separately.
Deployment comes in three shapes: regional cloud endpoints, on-premise, and on-device. Inference stays in-region, which suits organisations with data residency obligations. Cartesia ranks first on both the Artificial Analysis speech and speech-to-text leaderboards and holds SOC 2 Type II certification.
Key Features
- Managed Agents: An end-to-end platform for phone voice agents, with numbers, concurrency and call management included so inbound and outbound calling works on arrival.
- Sonic synthesis: Ultra-realistic text to speech billed per character, with instant and professional voice cloning.
- Ink transcription: Low-latency streaming speech to text billed per second of audio, sharing an account with synthesis while keeping a separate concurrency pool.
- State space architecture: SSM and Mamba lineage models keep latency low across long sequences, which suits continuous conversation rather than sentence-by-sentence requests.
- Three deployment shapes: Cloud, on-premise and on-device run the same models, so inference can stay within a chosen region.
- Independent concurrency: Synthesis, transcription and agents each get their own concurrency allowance and can run at once without competing.
Use Cases
- Customer service lets an agent answer inbound calls, resolve routine questions and hand off the rest.
- Outbound reminders and follow-ups run at a per-minute cost that is easy to forecast.
- Product voice assistants move to lower-latency synthesis to cut dead air in conversation.
- Live captions and meeting transcription share one account with synthesis.
- Regulated organisations deploy on-premise to keep inference in-region.
- Device-side interaction works without a round trip to the cloud.
Pros
- Ranks first on both the Artificial Analysis speech and speech-to-text leaderboards.
- The SSM architecture keeps latency low for continuous dialogue.
- Agent calls bill per minute, so cost scales linearly with volume.
- One API covers synthesis, transcription and agents, avoiding multi-vendor glue.
- Cloud, on-premise and on-device options support data residency needs.
- SOC 2 Type II certified, with DPAs and BAAs available on Enterprise.
Pricing
Free is 0 dollars a month with 20K credits and 1 dollar of prepaid agent usage, about 33 minutes of calls. Pro is 5 dollars a month with 100K credits and 5 dollars of agent usage, adding a commercial licence and instant cloning. Startup is 49 dollars a month with 1.25M credits and 49 dollars of agent usage, adding professional cloning and organisations, roughly 27 hours of calls. Scale is 299 dollars a month with 8M credits and 299 dollars of agent usage, adding priority support and high concurrency, about 166 hours of calls. Enterprise is quoted. Sonic synthesis costs 1 credit per character, roughly 750 to 800 credits per minute of audio, and Ink transcription costs 3 credits per second of audio. Agent calls are 0.03 dollars per minute with 0.014 dollars per minute of telephony on a Cartesia number, and any LLM used during the call bills separately. Overage on model credits runs 65 dollars per million on Pro, 45 on Startup and 38 on Scale.
Summary
Voice agents have an unusual cost structure: a slightly weaker model goes unnoticed, but a slightly higher latency makes a conversation fall apart. Cartesia bets on architecture, using state space models to keep latency down, then stacks synthesis, transcription and a calling platform on top so enterprises deal with one vendor. Per-minute billing also makes it possible to say what a single outbound call actually cost.
Version History
- Cartesia listed: Managed Agents voice platform (2026-10-31): Form at listing: a real-time voice platform built on in-house state space models, split into Sonic synthesis, Ink streaming transcription and Managed Agents for phone voice agents. Managed Agents includes numbers, concurrency and call management, billing 0.03 dollars per minute for calls, 0.014 dollars per minute for telephony and separately for any LLM used. Cloud, on-premise and on-device deployment with in-region inference, and independent concurrency pools for synthesis, transcription and agents. Free at 20K credits, Pro 5 dollars, Startup 49 and Scale 299, Enterprise quoted. Ranks first on both Artificial Analysis voice leaderboards, SOC 2 Type II.