Cartesia public-data review: what the credits actually buy
Cartesia is the cheapest text-to-speech option listed on this site by a wide margin: Pro is $5 a month for 100,000 credits, which Cartesia advertises as about 133 TTS minutes — roughly $0.038 per minute, against about $0.18 on ElevenLabs Starter. The catch is not the price, it is the pool. Text to speech and speech to text draw from the same monthly credits, and which transcription model you call changes your STT hours by 3x at identical cost. Voice agents are metered separately in dollars, not credits.
The plans, as Cartesia publishes them
Cartesia sells three products from one account: Sonic for text to speech, Ink for streaming speech to text, and Managed Agents for production voice agents. The company describes its models as built on state space models rather than transformers, which is the technical basis for its latency pitch.
| Plan | Price | Credits / month | Advertised TTS minutes | Advertised STT hours |
|---|---|---|---|---|
| Free | $0 | 20K | ~27 min | ~1h 51m |
| Pro | $5/mo | 100K | ~133 min | ~9h 16m |
| Startup | $49/mo | 1.25M | ~1,667 min | ~115h 44m |
| Scale | $299/mo | 8M | ~10,667 min | ~740h 44m |
| Enterprise | Custom | Custom | Custom | Custom |
Every tier includes unlimited workspace seats, which is unusual in this category — most rivals charge per editor seat. Each plan also carries a prepaid agents allowance denominated in dollars rather than credits ($1 on Free, $5 on Pro, $49 on Startup, $299 on Scale). Commercial-use licensing and instant voice cloning start at Pro; professional voice cloning and organisations start at Startup.
One thing to note about the sticker price: several third-party pricing roundups list Cartesia at $4, $39 and $239 for Pro, Startup and Scale. Cartesia's own pricing page showed no monthly/annual toggle during this review and listed $5, $49 and $299, so those lower figures are most likely annual-billing rates. They are not quoted from a Cartesia page here, so treat them as unconfirmed and check the checkout screen.
What a credit actually buys
Cartesia's docs pricing page publishes the underlying rates, and they are simpler than most credit systems on this site — there is no per-avatar or per-model multiplier hiding in the middle.
| Operation | Official credit rate |
|---|---|
| Text to speech (Sonic) | ~1 credit per character |
| Speech to text, ink-2 | 3 credits per second of audio |
| Speech to text, ink-whisper (websocket) | 1 credit per second of audio |
| Speech to text, ink-whisper (batch endpoint) | 1 credit per 2 seconds of audio |
| Infill | Fixed 300 credits + standard TTS rate on the infill transcript |
| Added voice accent | 225 credits |
| Voice agent call | $0.06 per minute (not credits) |
| Telephony, Cartesia-provided number | +$0.014 per minute |
Failed requests are free: Cartesia's docs state that errors will not consume credits.
The 750-characters-a-minute assumption
Divide each plan's credits by its advertised TTS minutes and you get the same number every time: 750 credits per minute. On Pro that is 100,000 ÷ 133; on Startup 1.25M ÷ 1,667; on Scale 8M ÷ 10,667. Since text to speech costs about 1 credit per character, Cartesia's minute estimates assume a script density of roughly 750 characters per minute of speech. That arithmetic is CartSignal's, applied to Cartesia's own published figures.
That assumption is on the generous side. Ordinary narration runs closer to 800–900 characters per minute once spaces and punctuation are counted. At 850 characters a minute, Pro's 100,000 credits deliver about 118 minutes rather than the advertised 133 — roughly 12 percent fewer. It is not a large gap, but it is the right direction to plan in: the advertised minutes are a modelled estimate, not an allowance you are guaranteed.
The transcription model that triples your hours
This is the biggest cost lever on the platform and it is easy to miss. Cartesia's advertised STT hours line up exactly with the ink-2 rate of 3 credits per second — 10,800 credits an hour. Pro's 100,000 credits ÷ 10,800 is 9.26 hours, which is the "~9h 16m" Cartesia prints. The same check holds on all four tiers.
But ink-whisper over websocket costs 1 credit per second, and its batch endpoint costs 1 credit per 2 seconds. For the same credits and no change of plan, that is roughly 3x the transcription hours streaming, or 6x batched. Pro's 100,000 credits go from about 9 hours on ink-2 to about 28 hours on ink-whisper streaming, or about 56 hours batched. Cartesia advertises the conservative number. If transcription volume is your constraint rather than accuracy, the model you pick matters far more than the tier you buy.
Cost per finished minute
| Plan | Price | Advertised TTS minutes | Cost per TTS minute | Cost per million characters |
|---|---|---|---|---|
| Pro | $5/mo | ~133 | ~$0.038 | ~$50 |
| Startup | $49/mo | ~1,667 | ~$0.029 | ~$39 |
| Scale | $299/mo | ~10,667 | ~$0.028 | ~$37 |
The volume discount is real but shallow — moving from $5 a month to $299 a month cuts the per-minute rate by only about a quarter. Unlike HeyGen, where the higher tier costs more per credit, Cartesia's curve at least points the right way. Buy the tier that matches your volume, not the one that looks like a bargain.
Concurrency is the real ceiling on the cheap tiers
Credits are not the only limit, and for anything real-time they are usually not the binding one. Cartesia's concurrency docs publish hard caps per plan.
| Plan | Concurrent TTS requests | Concurrent STT requests |
|---|---|---|
| Free | 2 | 8 |
| Pro | 3 | 12 |
| Startup | 5 | 20 |
| Scale | 15 | 60 |
| Enterprise | Custom | Custom |
Three concurrent TTS requests on a $5 plan is fine for a batch narration script and not enough for a live product with simultaneous users. Exceeding the cap returns a 429 Too Many Requests. TTS and STT limits are tracked separately and do not count against each other, and websocket connections are capped at 10x your concurrency limit — 150 on Scale. Idle websockets are closed after 5 minutes for TTS and 3 minutes for STT, so long-lived connections need reopening.
Voice cloning, and a figure the internet gets wrong
Instant voice cloning starts at Pro; professional voice cloning (PVC) starts at Startup. Cartesia's professional voice cloning announcement states that PVC training is free and that generating PVC speech costs 1 credit per character, the same as instant voice cloning, with clones available self-serve rather than through sales.
That matters because a number of third-party pricing roundups still state that PVC costs 1.5 credits per character with a one-time training fee of 1,000,000 credits. Cartesia's own announcement contradicts both figures. This review could not retrieve the current cloning rate from Cartesia's authenticated docs to settle it definitively, so the honest position is: the vendor's own published statement says free training at the standard rate, the widely-repeated 1.5x figure appears to be outdated, and you should confirm at signup if cloning volume drives your budget.
Overage behaviour
Cartesia's docs describe an overage switch. With overages disabled, requests that would exceed your allotment simply fail until the next renewal or until you upgrade. With overages enabled, the extra usage is billed on top of your subscription. That is a meaningful default to check before launching anything customer-facing — the safe-for-budget setting is the one that breaks your product mid-month.
Whether unused credits carry forward is the one commercial question this review could not answer from an official page. Cartesia's pricing FAQ lists "What happens to my rollover credits if I change my pricing tier?", which implies rollover exists in some form, but the answer was not retrievable and the docs pricing page states no expiry rule. Several third-party roundups assert credits expire at the end of the cycle. It is left unresolved here rather than guessed.
What Cartesia actually does
- Sonic-3.6 text to speech. Cartesia advertises "sub-90ms latency" across 44 languages, with emotion and tone that adapt to conversational context.
- Voice cloning from 10 seconds of audio, with localisation into other languages while preserving speaker identity.
- Ink speech to text. Two models — ink-2 and ink-whisper — at materially different credit rates.
- Managed Agents. Production voice agents with telephony, billed per minute in dollars.
- Pronunciation dictionaries for specialist terminology, and non-verbal expressions such as laughter embedded directly in the transcript.
- Unlimited workspace seats on every tier, including Free.
Cartesia's marketing claims a #1 ranking on Artificial Analysis's Speech Arena and speech-to-text leaderboards, and names ServiceNow, Retell and Quora among its customers. Those are vendor-published positioning claims reproduced here, not results CartSignal has verified.
Best-fit use cases
- Voice agents and live conversational products, where time-to-first-audio is the deciding spec
- High-volume narration on a tight budget — this is the cheapest per-minute TTS listed on this site
- Developers who want speech generation and transcription billed from one pool and one API
- Teams that need many collaborators, given seats are unlimited on every tier
- Transcription-heavy workloads, provided you pick the model that matches your accuracy needs
Limitations to verify
- Credit rollover and expiry are not documented on an official page retrievable during this review
- Concurrency, not credits, is the practical ceiling below Scale — 3 concurrent TTS requests on Pro
- Advertised TTS minutes assume ~750 characters per minute; denser scripts deliver fewer
- Advertised STT hours assume the more expensive ink-2 model
- The overage switch either fails your requests or bills you — confirm which default you are on
- Commercial-use licensing starts at Pro, so the free plan is not usable for client work
- Professional voice cloning requires the $49 Startup tier or above
How it fits against other tools here
Cartesia is the reason the ElevenLabs vs PlayHT comparison still has a cheap option to recommend. PlayHT's pitch was "most of the quality, less of the ElevenLabs bill", and after that product shut down, Cartesia is the listing that occupies the same slot on price — about $0.038 per TTS minute on Pro against roughly $0.18 on ElevenLabs Starter. If you chose PlayHT because it was cheaper and you were calling an API, Cartesia is the closest replacement; the PlayHT page covers what happened and what the migration involves.
The comparison to be careful with is quality-for-price. ElevenLabs competes on voice realism, dubbing and cloning depth, and its credit breakdown shows how differently each of its products burns the same pool. Cartesia competes on latency and throughput. For a podcast intro the price gap probably decides it; for a flagship brand voice it probably does not. Murf AI's Falcon engine is the direct competitor to Cartesia's real-time pitch, and Murf is the better pick if you want a timeline editor rather than an API. Resemble AI is the one built around watermarking and governance, which matters if you are working through the disclosure duties in our EU AI Act labeling guide.
The wider lists are best AI voice tools and the ElevenLabs alternatives shortlist, and everything in the category sits on the AI voice tools page.
FAQ
How much does Cartesia cost?
As of September 2026: Free at $0 with 20,000 credits a month, Pro at $5/month with 100,000, Startup at $49/month with 1.25M, Scale at $299/month with 8M, and a custom Enterprise tier. All plans include unlimited workspace seats. Voice agent calls are metered separately at $0.06 per minute, plus $0.014 per minute for a Cartesia-provided phone number.
How many minutes of speech does a Cartesia credit buy?
Text to speech costs about 1 credit per character, and Cartesia's advertised minutes work out to a flat 750 credits per minute on every tier — implying a script density of roughly 750 characters per minute. Narration often runs denser, so at about 850 characters a minute Pro delivers roughly 118 minutes rather than the advertised 133.
Is Cartesia cheaper than ElevenLabs?
On list prices, yes — about $0.038 per TTS minute on Cartesia Pro against roughly $0.18 on ElevenLabs Starter. Corrected 12 September 2026: this page previously put that gap at 4.8x, but the two per-minute figures were derived at different speaking densities (750 characters per minute for Cartesia, 900 for ElevenLabs), which inflated it. Per character — the unit that needs no density assumption — Cartesia Pro is $50.00 per million against ElevenLabs Starter's $200.00, so the real gap is exactly 4x. The full field ranked in that unit is in ElevenLabs alternatives. That is a price comparison only. The two are positioned differently: Cartesia on latency and throughput, ElevenLabs on voice realism, dubbing and cloning.
Do Cartesia credits roll over or expire?
Not resolved from an official page during this review. Cartesia's pricing FAQ implies rollover exists by asking what happens to rollover credits when you change tier, but the answer was not retrievable and the docs state no expiry rule, while third-party roundups claim credits expire at cycle end. Confirm with Cartesia before pre-buying a large allowance.
Has CartSignal tested Cartesia?
No. This is a public-data review built from Cartesia's official pricing page, docs and product pages checked in September 2026. No hands-on latency, audio quality or transcription accuracy testing is claimed, and Cartesia's speed and ranking claims are reproduced as vendor claims.
Source links
- Official Cartesia pricing page (plans, credits, advertised minutes and hours, agent rates)
- Official Cartesia docs pricing (credit rates per operation, overage behaviour)
- Official concurrency limits and websocket timeouts
- Official Sonic model page (Sonic-3.6, latency and language claims)
- Official professional voice cloning announcement (free training, 1 credit per character)
- Official Cartesia homepage (product line and positioning)