Cartesia sells speech models and managed voice agents, with different meters for each. A useful budget separates generated speech, transcribed audio, agent call duration and phone service. For technical teams, the monthly plan price is only the starting point: the same workload can fit its credit allowance while exceeding a concurrency limit or its prepaid agent funds.
Cartesia pricing at a glance
Cartesia has a Free plan and three paid self-serve plans: Pro at $5/month, Startup at $49/month and Scale at $299/month, plus custom Enterprise pricing. Each plan includes two separate allowances: model credits and prepaid dollars for Managed Agents.
Text-to-speech (TTS) generates audio. Speech-to-text (STT) transcribes audio. Direct use of those model APIs consumes credits. Managed Agent calls use a separate dollar meter. Telephony and language-model costs are separate.
The Managed Agent runtime rate is $0.03/minute on Free through Scale, with custom Enterprise pricing.
The plan table separates the monthly allowances:
| Plan | Monthly subscription | Monthly model credits | Included agent dollars/month |
|---|---|---|---|
| Free | $0 | 20,000 | $1 |
| Pro | $5 | 100,000 | $5 |
| Startup | $49 | 1,250,000 | $49 |
| Scale | $299 | 8,000,000 | $299 |
| Enterprise | Custom | Custom | Custom |
The agent allowance is bundled into the subscription. Startup costs $49/month and includes $49 of agent usage alongside its model credits. It does not cost $49 for the plan plus another $49 to receive that allowance.
Keep those balances separate in a forecast. A spare model-credit balance is not an additional dollar allowance for agent calls. For budgets spanning multiple providers, our voice-agent pricing guide explains how runtime, model and carrier charges fit together.
Speech-model credits: generated output and transcribed input
Sonic TTS uses approximately one credit per character. Transcript preprocessing can change the exact count slightly. The rate applies across /tts/bytes, /tts/sse and /tts/websocket.
Cartesia estimates 750–800 credits per minute of generated speech. That is a planning range, not a fixed conversion from call minutes. A five-minute conversation does not necessarily contain five minutes of agent speech.
For STT, model and endpoint both matter. The usage pricing rules distinguish realtime transcription from batch processing:
| Model | Realtime STT | Batch /stt |
|---|---|---|
ink-2 | 3 credits per audio second | Not available |
ink-whisper | 1 credit per audio second | 1 credit per 2 audio seconds |
The realtime rates apply to /stt/websocket and /stt/turns/websocket. Silence in submitted audio still counts, even when it produces no transcript. Ink 2 therefore consumes 180 credits per input minute, or 10,800 per input hour.
TTS and STT draw from the same monthly model-credit pool. Free's 20,000 credits represent roughly 27 minutes of Sonic output or 1 hour 51 minutes of Ink 2 input. Those are alternative uses of the pool. Using both reduces the amount available for each.
This creates three different durations in a custom voice stack:
- Generated speech duration: the audio your TTS requests produce.
- Transcription input duration: all audio submitted to STT, including silence.
- Call duration: the connected conversation, which may contain listening, speaking and waiting.
Treating all three as the same number can distort the estimate before any dollar conversion.
Other model operations have their own rules. Audio infill, which generates a bridge between existing clips, costs 300 credits per request plus the standard TTS credits for its transcript. Successful requests consume credits; errors do not.
Model-credit overages and rollover
The billing FAQ lists optional model-credit overages for paid plans. The additional rates are $65 per million credits on Pro, $45 on Startup and $38 on Scale. The rate applies to excess credits, rather than all credits consumed that month.
With overages enabled, requests can continue beyond the allowance and the credit balance can go negative. With overages disabled, a request that would exceed the available balance fails. Buying more monthly capacity and allowing usage to continue past a balance are separate decisions.
Paid credits remain when changing tiers. The rollover cap becomes twice the new plan's monthly credit allowance. An upgrade adds the new tier's credit amount to the existing balance. Cancellation or downgrade takes effect at the end of the billing period; existing credits remain until used. These are model-credit rules, separate from the monthly prepaid agent-dollar allowance.
Managed Agents: runtime funds, language models and phone service
Managed Agents bill call duration in U.S. dollars. Runtime charges do not draw from the model-credit balance. Once the included agent dollars are exhausted, additional agent usage continues at the plan's agent rate.
Dividing each plan's included agent dollars by the $0.03/minute runtime rate gives these runtime-minute estimates, rounded to the nearest minute:
| Plan | Included runtime at $0.03/min |
|---|---|
| Free | About 33 minutes |
| Pro | About 167 minutes |
| Startup | About 1,633 minutes |
| Scale | About 9,967 minutes |
These figures cover runtime alone, excluding phone service and language-model tokens.
A large language model (LLM) generates the agent's responses. Its charge depends on the selected model and token usage. Cartesia's authenticated GET /v1/agents/models catalog returns account-available models and token prices, including input, output and cache pricing.
The managed-model documentation describes that account-specific catalog. There is no single LLM surcharge that can accurately be applied to every Cartesia agent minute.
LLM usage is billed separately. A budget that excludes LLM charges is a subtotal.
A free number still has a telephony meter
Cartesia-provisioned numbers have no number rental charge. Calls using them add $0.014/minute to the agent rate, for a combined runtime-and-telephony subtotal of $0.044/minute ($0.03 + $0.014).
That subtotal excludes LLM charges and precedes any offset from included agent funds.
Session Initiation Protocol (SIP) connects a voice platform to a phone carrier. The phone connection options determine who bills the carrier portion:
| Phone connection | Cartesia telephony add-on | Geographic scope |
|---|---|---|
| Cartesia-provisioned number | $0.014/min | U.S. numbers; outbound to U.S. destinations only |
| Imported Twilio number | None; Twilio charges apply | Twilio-supported countries and routes |
| SIP trunk number | None; carrier charges apply | Carrier-dependent |
Bringing a Twilio number or SIP trunk moves the phone-service bill to that provider. It does not remove the carrier cost or Cartesia's agent-runtime charge.
Imported Twilio and SIP numbers do not consume Cartesia-provisioned number slots. A phone-number slot and a concurrent call slot are also different limits: one concerns numbers assigned to the account, the other simultaneous conversations.
Plan limits: monthly volume is not peak capacity
Concurrency means work running at the same time. Cartesia maintains separate pools for TTS requests, STT requests and Managed Agent calls. Activity in one pool does not consume capacity in another.
| Plan | TTS requests | STT requests | Agent calls | Cartesia numbers |
|---|---|---|---|---|
| Free | 2 | 8 | 8 | 1 |
| Pro | 3 | 12 | 12 | 3 |
| Startup | 5 | 20 | 20 | 5 |
| Scale | 15 | 60 | 60 | 10 |
| Enterprise | Custom | Custom | Custom | Custom |
The first three numeric columns are simultaneous-work limits. The final column is the number of Cartesia-provisioned numbers.
For example, Startup's five direct TTS requests and 20 agent calls are independent limits. It would be incorrect to describe Startup as allowing only five Managed Agent calls because its TTS API limit is five. Equally, its agent-call capacity does not entitle a custom stack to 20 simultaneous direct TTS requests.
Credits measure monthly volume. Concurrency governs peaks. A bursty workload can fit comfortably inside the monthly credit pool and still exceed its direct API concurrency limit.
Feature requirements can also determine the tier:
- Pro adds a commercial-use license and instant voice cloning.
- Startup adds Organizations and two Professional Voice Clone slots per organization.
- Scale adds priority support and four Professional Voice Clone slots per organization.
- Enterprise has custom usage and concurrency terms.
Training a Professional Voice Clone is free. Speech generated with it uses one credit per character unless an Enterprise agreement changes that rate. Voice localization costs 225 credits per added accent. A cloning slot is a capacity entitlement, rather than free unlimited speech generation.
Two small workloads, without mixing balances
These illustrative planning scenarios use the stated model-credit and agent-runtime rates, with no other account usage during the month. They are separate workloads, not two charges to add to the same calls.
Workload A: direct speech APIs
Assume 15 minutes of generated Sonic speech and 15 minutes of total Ink 2 realtime input, including silence. There are no Managed Agent calls or phone charges in this scenario.
| Usage | Calculation | Model credits |
|---|---|---|
| Sonic output | 15 minutes × 750–800 credits | Approximately 11,250–12,000 |
| Ink 2 input | 900 seconds × 3 credits | 2,700 |
| Combined usage | TTS + STT | Approximately 13,950–14,700 |
This fits within Free's 20,000 monthly credits, leaving approximately 5,300–6,050 credits. For this isolated model-API workload, the plan fee and model-credit overage are both $0.
That result establishes a volume budget only. The workload must also stay within Free's two simultaneous TTS requests and eight simultaneous STT requests. The $0 estimate covers usage only; the commercial-use license is a Pro-and-above feature.
Workload B: Managed Agent phone calls
Assume ten sequential calls lasting exactly five minutes each, using one Cartesia-provisioned number on Free. That is 50 agent minutes and 50 telephony minutes. There are no separate direct TTS or STT API requests in this scenario.
The runtime-and-telephony subtotal applies Free's included agent funds to runtime only:
| Charge | Calculation | Amount |
|---|---|---|
| Gross agent-runtime usage | 50 × $0.03 | $1.50 |
| Included agent funds | Applied to runtime | −$1.00 |
| Additional agent charge | $1.50 − $1.00 | $0.50 |
| Telephony | 50 × $0.014 | $0.70 |
| Subscription | Free plan | $0.00 |
| Runtime + telephony subtotal | $0.50 + $0.70 | $1.20 |
One number and one simultaneous call fit Free's published limits. The $1.20 runtime-and-telephony subtotal is not an all-in bill or contractual quote. Model-specific LLM charges, taxes and other account usage are excluded. Whether the agent's built-in speech processing also incurs a separate model-credit debit is unspecified; any such debit is excluded from the subtotal.
The example uses whole five-minute calls and makes no assumption about Cartesia's billing increments for partial minutes. The direct model-API credits from Workload A are not added to Workload B.
Build the budget around the architecture
For a team using Sonic or Ink inside its own stack, the budget starts with model credits. It also includes the runtime, LLM, carrier and operating infrastructure supplied elsewhere. Generated speech and submitted transcription audio remain separate inputs.
For Managed Agents, the budget starts with the subscription and agent dollars, then phone service and applicable LLM usage. Included agent dollars offset eligible runtime usage once. They are not a second subscription charge or a reason to omit the phone bill.
A plan decision therefore has three constraints: monthly consumption, peak concurrency and required features. None can stand in for the other two.
When a managed production platform is the purchase
If you need only a speech model in an existing application, Cartesia's direct API credit model is the relevant evaluation. If you are choosing the foundation for a production conversational AI product, the runtime and operating workflows belong in the decision too.
At Dasha, we help technical teams build and run production voice AI agents through a managed runtime, REST APIs and a web application, with telephony, integrations, testing and monitoring.
Our Growth pricing starts at $0.08/minute, excluding Voice over Internet Protocol (VoIP) and LLM tokens. We bill in per-second increments with no minimum call duration and no charge for unconnected call attempts. That is a different bundle from Cartesia's individual rate lines; the headline rates alone do not establish which deployment costs less.
For a managed production foundation, start building with Dasha.
