A voice agent that runs out of its monthly allowance mid-afternoon has a billing problem that callers hear. Cartesia sells text to speech as monthly plans of credits, one credit per character, and says that with overages switched off "new requests are blocked once your credits run out." ElevenLabs bills its API in US dollars per character, which sounds like a small difference until a team sizes a plan for a real call volume.
Speed is not what separates these two, since both publish first-audio figures near a tenth of a second. What does separate them is billing, which models run on which endpoint, and how the voices scored with listeners, and Bland Speech belongs beside them for a team building phone agents.
Which is better, ElevenLabs or Cartesia?
As of November 2026, ElevenLabs is the better fit for a team that wants the widest language coverage and a cloning ladder that starts on a $6 plan, and Cartesia is the better fit for a team that wants the lower published first-audio claim and a credit that equals one character. For a phone line, a third option, Bland Speech, is worth pricing next to both.
What these rows mean
- Billing and rate: Bland Speech and ElevenLabs charge per character of text sent. Cartesia sells a monthly pot of credits and charges overage rates only when a team switches overages on. ElevenLabs prices its API separately from its monthly plans.
- First audio as published: The time from sending text to hearing the first sound. Each number is what that company publishes, measured its own way.
- Mu-law and A-law: Two compressed audio formats used on phone networks, mu-law in North America and Japan and A-law in most other countries. Both are 8-bit audio at 8 kHz.
- Realism result: The Elo score from Design Arena's Audio Realism Bench results of 4 August 2026, where listeners pick which of two clips sounds more like a real person. Eleven v4 and Sonic 3.6 were released after those results and weren't scored in them.
- Instant and professional clone: An instant clone is built from a short sample, a professional clone is trained on a long recording of one speaker.
What does ElevenLabs do well, and where does it fall short on a call?
ElevenLabs does well on range and falls short on where its newest models can run. ElevenLabs calls Eleven v4 Turbo its state of the art model for real-time speech and gives a median latency of about 100ms, excluding application and network latency, and ElevenLabs says the Eleven v4 family supports 90+ languages. Audio tags such as [whispering], [shouting], and [laughing] work on Eleven v4, which gives a developer a way to direct a line without rewriting it.
The catch sits in the plumbing. ElevenLabs says its standard text to speech WebSocket "does not support the eleven_v3 or eleven_v4 models", so Eleven v4 runs through the Text to Dialogue API and Eleven v4 Turbo through the Text to Dialogue WebSocket. A team that wired an agent to the standard WebSocket on Flash v2.5 has a migration to plan before it can use the new models. ElevenLabs also gives two figures for one model, about 100ms and about 75ms for Flash v2.5.
ElevenLabs pricing
ElevenLabs charges API usage in dollars: $0.08 per 1,000 characters on Eleven v4, $0.04 on Eleven v4 Turbo and the Flash models. Plans run Starter at $6 a month, Creator at $22, Pro at $99, Scale at $299, and Business at $990, with Enterprise custom. At the v4 rate those plans include 75,000, 275,000, 1,238,000, 3,738,000, and 12,375,000 characters, and the figures double at the v4 Turbo and Flash rate.
ElevenLabs counts credits on its creator plans and dollars per character on its API plans. A free account gets 10,000 characters at the v4 rate and is limited to non-commercial use, and paid plans carry the commercial licence.
ElevenLabs cloning and compliance
An instant clone needs 10 seconds of audio on Eleven v4 and starts on the Starter plan. A professional clone needs about 30 minutes, a Creator plan or above, and a voice-captcha check where the voice owner reads lines aloud to confirm permission. ElevenLabs offers Business Associate Agreements to qualifying healthcare enterprises and says it holds SOC 2 Type 2 and that teams can run text to speech inside their own cloud infrastructure.
Pros
- 90+ languages on the Eleven v4 family
- Audio tags that direct whispering, shouting, and laughing
- Cloning from a $6 plan, with a captcha-style permission check
- Mu-law and A-law output at 8 kHz for phone lines
Cons
- Eleven v3 and v4 don't run on the standard text to speech WebSocket
- Eleven v4 costs $0.08 per 1,000 characters, twice the Turbo and Flash rate
- 44.1 kHz PCM output needs the $99 Pro plan
- Flash v2.5 latency given as both about 100ms and about 75ms
ElevenLabs is a poor fit for a live agent that has to stay on the standard WebSocket, and for a team that wants Eleven v4 at the Turbo price.
What does Cartesia do well, and where does it fall short on a call?
Cartesia does well on speed and developer controls and falls short on how credits meter a busy line. Cartesia says Sonic delivers first audio in under 90ms and sells it with three streaming endpoints, bytes, SSE, and WebSocket, where only the WebSocket takes text that arrives in pieces from a language model. Word-level timestamps come back on the WebSocket and SSE endpoints, pronunciation dictionaries take IPA or "sounds-like" entries by id, and Sonic 3.6 speaks 44 languages.
The catch is the meter. Cartesia counts one credit per character and about 750 to 800 credits per minute of audio, so the $5 Pro plan's 100,000 credits cover about 133 minutes. Each break tag costs a credit, and adding an accent to a voice costs 225 credits.
Cartesia pricing
Cartesia's plans are Free at $0 with 20,000 credits, Pro at $5 with 100,000, Startup at $49 with 1.25 million, Scale at $299 with 8 million, and Enterprise custom. Overages are opt-in at $65, $45, or $38 per million credits on Pro, Startup, or Scale, billed when the plan renews, and unused credits roll over up to twice the monthly allotment.
Cartesia cloning and compliance
An instant clone needs 10 seconds of audio, up to 60 for a better accent match. A professional clone needs the Startup plan, training takes up to 3 hours, and Cartesia doesn't charge for the training. Speed and volume controls don't apply to a professional clone, because the clone learns pacing and loudness from its recordings. Cartesia is SOC 2 Type II certified, with BAAs and DPAs on Enterprise, and says Sonic can run on-premise in an air-gapped environment, in a customer's own VPC, or under an OEM licence.
Pros
- First audio in under 90ms, as Cartesia publishes it
- Three streaming endpoints, with word timestamps on two of them
- 44 languages and pronunciation dictionaries
- Professional clone training is free
Cons
- With overages off, requests stop at zero credits
- Professional cloning starts on the $49 Startup plan
- Break tags and added accents cost credits
- Emotion controls are beta and English only
Cartesia is a poor fit for a line that must never stop at an allowance. A developer who built agents on Cartesia's Line SDK also has a deadline, because Cartesia stops hosting Line SDK agents on January 1, 2027 and points teams to Managed Agents.
Which sounds more human, ElevenLabs or Cartesia?
On Design Arena's Audio Realism Bench results of 4 August 2026, Cartesia's Sonic 3.5 ranked 6th of 13 AI voices at 1096 and ElevenLabs' Eleven v3 ranked 12th at 938, so Cartesia's model scored 158 points higher. Elo is a rating that rises each time listeners pick a voice as more human, and Design Arena describes the test as "two recordings of the same transcript, (presented blindly and in randomized order)" judged by vetted native speakers.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 3rdxAIGrok TTS1156
- 4thOpenAIGPT Realtime 21119
- 5thGoogleGemini 2.5 Pro TTS1109
- 12thElevenLabsEleven v3938
Ranks are among 13 AI voices, and ElevenLabs's own voice is on the list. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
Neither company's current model was in those results. Cartesia released Sonic 3.6 on 27 August 2026 and ElevenLabs released Eleven v4 and Eleven v4 Turbo on 28 September 2026, so the scores describe the models each company shipped before them. Bland Speech v3 ranked 1st at 1365 in the same results, and the human recording scored 1500. The method is in Design Arena's Audio Realism Bench article.
Why do people comparing ElevenLabs and Cartesia also weigh Bland Speech?
People comparing ElevenLabs and Cartesia for a phone agent also weigh Bland Speech because Bland Speech v3 ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026, at 1365, and bills one standard rate with no monthly plan. Bland Speech costs $0.04 per 1,000 characters, pay as you go, with 50,000 free characters once for a new account, so there is no monthly allowance to run into.
Bland publishes 199ms to first audio, which is higher than Cartesia's under-90ms and ElevenLabs' median of about 100ms, so a team that weighs first-audio claims above everything else will pick one of the rivals. The trade is realism and billing for roughly a tenth of a second, and on a phone call, where a caller hears the voice for minutes and the gap for a moment, the voice is worth more, so for phone calls and voice agents Bland Speech is the one to choose.
Bland Speech streams over an HTTP chunked endpoint and a WebSocket that accepts text as a language model writes it, and a new context_id interrupts the turn that is playing, which is how barge-in works when a caller talks over the agent. Output is PCM at 8 to 48 kHz or mu-law at 8 kHz. Performance tags such as [laughs] and [say warmly] shape a line. An instant clone needs one clean sample of about 10 seconds, a professional clone is fine-tuned on 30 minutes or more of audio and goes through Bland's consent and rights review, and every clone is private to the customer's organization. Bland Speech has a first-party Pipecat service and a LiveKit Agents plugin for Python, and the text to speech API page has the endpoints.
Bland is SOC 2 Type II audited, HIPAA compliant with a signed BAA, and FedRAMP 20x Class C certified at Moderate, granted 17 November 2026, and Bland's trust page says speech synthesis runs inside the FedRAMP boundary. Enterprise pricing is custom, set with Bland's sales team.
Pros
- Ranked 1st of 13 in the 4 August 2026 realism results
- One standard rate, $0.04 per 1,000 characters, with no monthly plan
- Barge-in built into the WebSocket protocol
- Pipecat and LiveKit integrations
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
The Bland Speech vs ElevenLabs page sets the two head to head, best Cartesia alternatives covers the other direction, and the text to speech hub holds every comparison, including the rival vs rival family.
Frequently asked questions
Is ElevenLabs better than Cartesia?
ElevenLabs is better for a team that needs 90+ languages and cloning from a $6 plan, and Cartesia is better for a team that wants the lower published first-audio claim of under 90ms against ElevenLabs' median of about 100ms. Neither company's current model was in Design Arena's 4 August 2026 results, so no score separates their newest models.
Which is cheaper, ElevenLabs or Cartesia?
Cost depends on volume and model. At ElevenLabs' API rates, 1 million characters costs $80 on Eleven v4 and $40 on Eleven v4 Turbo, and ElevenLabs estimates about 1,000 characters a minute. Cartesia estimates 750 to 800 credits a minute, so 1,000 minutes fits inside the $49 Startup plan's 1.25 million credits. Bland Speech is $40 per million characters, with no plan.
Does Cartesia or ElevenLabs support phone audio formats?
Both do. ElevenLabs offers mu-law and A-law at 8 kHz among its output formats, and Cartesia offers mu-law and A-law with sample rates from 8,000 to 48,000 Hz. Bland Speech offers mu-law at 8 kHz and PCM at 8,000 to 48,000 Hz.
Can ElevenLabs or Cartesia clone a voice from 10 seconds of audio?
Both can. ElevenLabs makes an instant clone from 10 seconds on Eleven v4 from the Starter plan, and Cartesia does the same, accepting up to 60 seconds for a better accent match. Bland Speech's instant clone needs one clean sample of about 10 seconds, and the customer confirms the voice is theirs to use.
Which has more languages, ElevenLabs or Cartesia?
ElevenLabs has more. ElevenLabs says its Eleven v4 family supports 90+ languages and Eleven v3 supports 70+, while Cartesia says Sonic 3.6 speaks 44 languages.
What happens when a Cartesia plan runs out of credits?
With overages off, new requests are blocked until credits renew. Overages are opt-in on paid plans and aren't available on the Free plan, and Cartesia bills them at renewal. Bland Speech bills each character as it's used from a wallet balance.
