Cartesia vs ElevenLabs: speed or the voice library?

Most of the versus pages on this site end up telling you the two products do different jobs. Not this one. Cartesia and ElevenLabs are the same kind of tool: each makes the voice itself, each will clone one from a sample, and each sells a stack for putting that voice on a live phone call. Same shelf, same buyer, a genuine head-to-head. And the fork between them is unusually clean. Cartesia is built around one promise, that speech starts in under 100 milliseconds, which is Cartesia’s own claim rather than anything we have measured, and around being cheap to run. ElevenLabs is built around the voice being the best available: the biggest library going, the best cloning, and output most people cannot tell from a human on a blind listen.

Quick map of where this goes. Price first, because for once the two bill in comparable units and the honest differences live in the footnotes. Then the speed question and why we will not referee it yet. Then who each one is built for, where each genuinely wins, the capability and compliance grid, a worked example, the bits we have not tested, and a straight answer at the end.

The price, told honestly

For a live phone agent, both charge by the minute, and the shapes are close enough to line up directly. Cartesia charges $0.06 a minute for the agent on its paid tiers, adds $0.014 a minute for a phone line on a Cartesia number, and leaves the AI model (the part that decides what to say) to you at roughly $0.01 a minute, so a realistic all-in lands $0.08 to 0.15. The listening half, Cartesia’s Ink-2 speech-to-text (the tech that turns the caller’s words into text), is bundled into the agent rate rather than billed separately. ElevenLabs charges $0.08 a minute for its premium voice, the same base rate on every plan tier, passes the AI model through at cost, and the phone line adds about $0.02 on Twilio, for a realistic all-in of $0.08 to 0.25. Same floor. Different ceiling. The gap is mostly the voice: you pay ElevenLabs a premium for the sound, and you pay Cartesia less for a faster, plainer engine.

For narration, the honest comparison needs a caveat before the numbers. ElevenLabs publishes clean per-character pricing on its subscriptions: Free gives 10,000 credits a month to try it, Starter is $6 for 30,000, Creator $22 for 121,000 (a first-month promo halves it), Pro $99 for 600,000, Scale $299 for 1.8 million and Business $990 for 6 million. One character is one credit on the standard models, which works out around $0.09 to 0.20 per 1,000 characters by tier and model, about $0.10 on the multilingual model most narration jobs use. Cartesia no longer publishes a per-character rate at all. It meters the same work in credits, about 15 credits per second of audio, so the roughly $0.035 per 1,000 characters we quote for it is a figure we derived by working the credits back, not a price printed on a page. Treat it as close rather than exact. Even with that caveat the gap is real: on our numbers, Cartesia narration runs at about a third of ElevenLabs’ standard rate.

One more Cartesia footnote, because it is fresh. When we re-captured its pricing page on 2026-07-11, the Pro tier had risen from $4 to $5 a month and Startup from $39 to $49, and the annual-billing toggle we recorded in June no longer shows. Small money at the Pro end, a quarter more at the Startup end, and worth knowing the direction of travel before you budget from an old blog post. The standing tiers are Free, Pro $5, Startup $49, Scale $299 and a custom Enterprise, with 100K, 1.25M and 8M credits a month included on the three paid self-serve tiers.

The speed question, and why we will not referee it

Cartesia’s whole pitch hangs on one number: speech that starts in under 100 milliseconds. In plain terms, that is the gap between the agent deciding what to say and the caller hearing the first sound of it, short enough that the reply lands inside the natural rhythm of a conversation instead of after that awkward beat you have heard on every bad phone bot. It is also, and this matters, Cartesia’s own published claim. We have not measured it.

ElevenLabs has a fast lane too. Flash v2.5 is the model it points real-time agents at, and its docs put it at about 75 milliseconds. Also a vendor figure. And here is the catch with lining the two up: they are not counting the same thing. Cartesia quotes time to first audio; ElevenLabs quotes the model’s generation time, before the phone line and the rest of the call stack add their share. Racing a 75 against a sub-100 tells you nothing when the stopwatches started at different points.

So this page will not referee the speed question, and honestly nobody else’s should either until someone runs the same stopwatch against both. That someone is eventually us: when our test harness ships we will place the same timed calls against both platforms and publish p50 and p95 with dates, and if the measured numbers contradict the marketing, the measured numbers win. Until then, every latency figure you see for either vendor, here or anywhere else, is the vendor marking its own homework.

What we can say without a stopwatch is where the live-agent market has voted. Cartesia’s own customers page reads like a directory of the call platforms we cover: Goodcall built its business phone agents on it, Vapi is named there as having chosen Cartesia as its default voice provider, and Retell appears on the same wall. Those are Cartesia’s own case studies, so apply the usual discount, but the pattern matches everything else on this page. Cartesia is the voice inside other people’s live agents, picked by teams whose product dies by the pause.

Who each one is built for

Two clean fits, and they sort most buyers:

• The call has to feel instant, and the bill has to stay predictable. Cartesia. You are building a live phone agent, probably wiring it yourself through Twilio, LiveKit or Pipecat, and the response gap is the thing your users will notice. You get speech-to-text bundled into a $0.06 agent rate and an all-in around $0.08 to 0.15 a minute. The trade sits in the same sentence: the tooling is aimed at developers rather than a click-around studio, and the surrounding call kit (warm transfer, batch dialling) has to come from your platform, because Cartesia does not carry it.
• The voice is what your audience judges. ElevenLabs. You are narrating videos or audiobooks, building a brand voice, or running an agent where callers will rate you on how it sounds. The 10,000-plus library, the cloning and the 70-plus languages do the heavy lifting. The trade, again in the same breath: you pay about three times Cartesia’s derived rate per character, and the compliance paperwork lives on the Enterprise plan.

Where Cartesia wins

Speed is the headline, so let us keep the label straight one more time: sub-100ms time to first audio is Cartesia’s claim, not our measurement. What we can credit it with is focus. The whole product is engineered around live, synchronous speech, the thing you buy when the pause after a caller stops talking is your enemy, and the customer wall of call platforms above suggests the people closest to that problem keep picking it.

The second win is the bill. Show the workings: $0.06 agent rate plus $0.014 telephony plus roughly $0.01 for your own AI model comes to about $0.08 a minute, and the realistic band tops out near $0.15. ElevenLabs starts at the same $0.08 floor but its band runs to $0.25. Over a month of real call volume, the ceiling is the number that hurts, and Cartesia’s is markedly lower. Speech-to-text being bundled also means one less line on the bill and one less vendor to reconcile.

The third win is narration cost. Roughly $0.035 per 1,000 characters against about $0.10 is a third of the price, and at content volume that compounds fast. The honest asterisks belong in the same sentence though: our Cartesia figure is worked back from credit metering rather than printed on its pricing page, and that pricing just moved, with Pro and Startup rising to $5 and $49 on the 2026-07-11 capture. Cheap today, derived today, and worth re-checking on the day you commit.

Cloning is quicker to start than you might expect too: an instant clone needs only a three-second clip, and the $5 Pro tier includes instant cloning with a commercial licence. But this is where the wins stop and the carve-outs take over. The library is around 100 voices across 42 languages (Cartesia no longer states a single public total), against ElevenLabs’ thousands. There is no warm transfer, no batch calling and no MCP support. And there are no compliance certificates in writing today: SOC 2 and HIPAA are advertised for the Enterprise tier, but we could not verify either from a primary source, so every compliance box on our grid stays unticked.

Where ElevenLabs wins

Voice quality first, with the label attached. Our editorial preview scores put ElevenLabs at 10 and Cartesia at 9, and those are provisional reads from public information and a first listen, not blind-test results. Cartesia sounding genuinely good is worth saying out loud. ElevenLabs being the one most people cannot tell from a human is why it holds our best-voice-quality badge, and if your output is judged on the sound alone, that last increment is the product.

Scale is the second win, and it is not close. The ElevenLabs library runs past 10,000 voices in 70-plus languages; Cartesia offers around 100 voices in 42 languages. Our range preview scores say the same thing, 10 against 8. If you need a specific accent, age or delivery, the odds ElevenLabs already has it are simply better, and the model line-up adds a second axis of choice: v3 and Multilingual v2 for rich, emotional narration, Flash v2.5 when speed matters.

Cloning depth is the third. Instant cloning for speed, professional cloning for a long-term brand voice, and it is widely regarded as the best at both. The proof points ElevenLabs publishes lean the same way: its customer stories feature Storytel rethinking audiobook narration and Headspace scaling meditation content across markets. Vendor-published stories, so read them as such, but they are the right kind of customer for the claim being made.

The fourth win is the ecosystem around the voice. ElevenLabs carries warm transfer (handing a live call to a human with the AI’s summary attached), batch calling for outbound campaigns, and MCP support, the Model Context Protocol connection that lets other AI tools trigger and feed your calls. Cartesia has none of those three. And when the buyer is regulated, ElevenLabs can put HIPAA, SOC 2 and GDPR in writing. The costs of all this sit in the same paragraph where they belong: the narration rate is about three times Cartesia’s derived figure, and those compliance guarantees live on the Enterprise plan, not the self-serve tiers.

Capability and compliance, side by side

The compliance row deserves its own paragraph, because for a chunk of buyers it ends the comparison. ElevenLabs puts HIPAA, SOC 2 and GDPR on its Enterprise plan, with EU data residency and a zero-retention mode available there; not cheap, not self-serve, but real and in writing. Cartesia advertises SOC 2 and HIPAA for its Enterprise tier, and we could not verify either from a primary certification source, so our vendor page leaves all four compliance boxes unticked until the paperwork shows. If you are in healthcare, finance or anywhere regulated, that is not a nitpick, it is the decision: get Cartesia’s certificates in writing before you build, or pick the vendor that already offers them.

A worked example, so the numbers feel real

Take a narration job first: 50 videos a month at about 8,000 characters each, so 400,000 characters. On ElevenLabs that sits comfortably inside the Pro plan at $99 a month for 600,000 credits, in the voice of your choice from a library of thousands. On Cartesia, 400,000 characters at the derived rate of $0.035 per 1,000 is about $14 of credits, and the $49 Startup tier’s 1.25 million monthly credits cover it several times over. Half the monthly bill, carrying the caveats you already know: the rate is derived, the tier prices just rose, and you are choosing from around 100 voices rather than 10,000. If none of those caveats bite, Cartesia wins this job on cost. If the read needs one specific voice that only a deep library has, ElevenLabs earns its premium.

Now a live agent: 5,000 minutes of inbound calls a month. Cartesia’s $0.08 to 0.15 band prices the month at roughly $400 to $750, with speech-to-text already in the rate. ElevenLabs’ $0.08 to 0.25 band prices it at roughly $400 to $1,250. Same floor, so a lightly used agent costs about the same either way; the ceilings diverge as volume and voice premium stack up. Which band you actually land in depends on your model choice and call mix, so run your own numbers through the cost calculator rather than trusting the midpoint of anyone’s range, ours included.

What we have not tested yet

The honest limits, gathered in one place. We have not placed a single timed test call to either platform, so there is no Voxrater latency number for Cartesia or ElevenLabs anywhere on this page; the sub-100ms and 75ms figures are the vendors’ own, counted in different ways, and we have labelled them as claims every time. The 1 to 10 scores are an editorial preview from public information and a first listen, not blind listening tests. When the harness ships we will run identical scenarios against both, publish p50, p95 and the dates, and let the measured numbers overrule the marketing wherever they disagree. Until then, this page is sourced and dated, but the speed verdict is deliberately missing, because we do not hand out verdicts on numbers we have not made.

Three questions that decide it

1. Is the pause the thing you are optimising? If the response gap is your product’s make-or-break, Cartesia is the shortlist, on the strength of a vendor claim we have not yet verified plus a customer wall of latency-obsessed call platforms. Our best low-latency roundup ranks the wider field.
2. Will your audience judge the voice itself? Narration, audiobooks, a brand voice, a premium agent: ElevenLabs, for the quality edge and the library depth. The narration rankings show how both stack up against the rest.
3. Do you need compliance in writing today? ElevenLabs can give it to you on Enterprise. Cartesia cannot show us a certificate yet, so regulated buyers either wait for the paperwork or look elsewhere.

Bottom line

Pick Cartesia when the job is a live phone agent and response speed decides it. Its entire product is built around the sub-100ms promise (Cartesia’s claim, ours to verify once the harness runs), the all-in cost of $0.08 to 0.15 a minute undercuts ElevenLabs’ ceiling by a wide margin, and the derived $0.035 per 1,000 characters makes it one of the cheapest narration engines we track. Accept the trades knowingly: about 100 voices in 42 languages, no warm transfer, batch calling or MCP, freshly risen tier prices, and no compliance certificate in writing today.

Pick ElevenLabs when the voice itself is what you are selling. It takes voice quality on our preview scores, the library is a hundred times the size, the language count runs 70-plus against 42, the cloning is the best around, and the operational kit plus Enterprise-plan compliance make it the safer institutional choice. Accept its trades too: roughly three times the narration cost per character, and the compliance conversation goes through sales, not a checkout page.

An even split, decided by your job, not by loyalty. Read the full Cartesia review and ElevenLabs review for the per-plan detail, then put your own call volume or character count through the cost calculator before you sign anything.