← Blog

ElevenLabs vs Cartesia: Which Voice AI API Fits Your Build?

November 19, 2026

AI CodingTool Comparisons

ElevenLabs vs Cartesia comes down to one question: are you making audio people listen to, or building a voice that talks back? Both companies now sell text-to-speech, speech-to-text and a voice agent platform, so the feature lists look alike. Where they differ is focus. ElevenLabs is a broad audio company with a huge voice library and expressive long-form models. Cartesia is built around speed for real-time conversation.

The short version, as of November 2026:

  • Pick ElevenLabs for narration, audiobooks, dubbing, characters and anything where expressive delivery matters, or if you need 90+ languages and a big ready-made voice library.
  • Pick Cartesia for real-time voice agents if you want a vendor built around speed (it claims first audio in under 90 ms), model snapshots pinned to a date, an on-premise option, and simple per-character billing.
  • Test both with your own scripts. Each vendor publishes benchmarks it wins. Your voice, your text and your network decide the result.

Model names in this space change every few months. Every name and status below was checked on each vendor's own docs on November 19, 2026.

ElevenLabs vs Cartesia: quick comparison

ElevenLabsCartesia
Flagship TTSEleven v4 (expressive), Eleven v4 Turbo (low latency)Sonic 3.6
Other current TTSEleven v3, v3 Conversational, Multilingual v2, Flash v2.5, Flash v2Older Sonic models still documented
Languages (flagship)90+ on Eleven v444 on Sonic 3.6
Speed claim, per vendorAbout 100 ms median inference for v4 Turbo and Flash v2.5First audio in under 90 ms
Instant voice replication10 seconds of audio, per the v4 launch post10 seconds of audio, per the Sonic page
Speech-to-textScribe v3 (batch), Scribe v3 Live (real time)Ink 2 (streaming), Ink preview
Voice agent platformElevenAgentsManaged Agents
TTS billing unitPer 1,000 characters, rate depends on the modelCredits, 1 credit per character
Self-hostingNot documented on the pages we checkedOn-premise or VPC through Enterprise sales

Current models, as of November 2026

ElevenLabs: Eleven v4 and v4 Turbo

ElevenLabs launched Eleven v4 on September 28, 2026, with a low-latency sibling, Eleven v4 Turbo. Both support more than 90 languages. The pitch for v4 is expressiveness: you can add inline audio tags such as [laughs] or describe in plain language how a line should be read, and ElevenLabs says v4 follows that direction more closely than earlier models.

The split is simple. Eleven v4 is for content: narration, long-form audio, dialogue. Eleven v4 Turbo is for agents and other interactive apps.

The ElevenLabs models page also lists Eleven v3, v3 Conversational, Multilingual v2, Flash v2.5 and Flash v2 as current. Two things older tutorials get wrong:

  • Turbo v2.5 and Turbo v2 are deprecated. Their replacements are Flash v2.5 and Flash v2. The page gives no removal date, but new code shouldn't use them.
  • Flash v2.5 is not deprecated. It's still the cheap, fast option, with 32 languages and the longest per-request character limit on the list.

Cartesia: Sonic 3.6

Cartesia's TTS models page calls Sonic 3.6 its fastest and most natural model, generally available, with 44 languages. Odia and Urdu are new in 3.6, and it handles Hinglish.

Cartesia's model IDs are worth understanding before you ship:

  • sonic-3.6 points to the latest stable snapshot and updates automatically.
  • A dated ID such as sonic-3.6-2026-08-27 never changes. Cartesia recommends dated snapshots for production.
  • sonic-preview is a beta that can change without notice.

If your app's sound needs to stay identical between releases, pin the dated snapshot. Professional voice replications made on older Sonic models still work on 3.6, per the same page.

Voice quality and expressiveness

This is where ElevenLabs has the stronger story for content work.

ElevenLabs gives you more ways to steer a performance. Audio tags, natural-language direction and a dialogue API for multi-speaker scripts are built into v4. It also has the larger ecosystem around the voice itself: a voice library of 10,000+ voices (per its agent docs), voice design, dubbing and sound effects. If you're producing a podcast intro, an audiobook chapter or a game character, that toolkit saves real time.

Cartesia takes a more automatic approach. Its Sonic page says Sonic reads the emotional subtext of a transcript and adjusts delivery by default, and that you can insert non-verbal sounds like [laughter] inline. That's enough for most conversational agents, where you want natural reactions without hand-directing every line.

On benchmarks, treat both sides as claims. ElevenLabs' v4 post reports winning most blind preference tests against a field that includes Cartesia Sonic 3.6, based on its own testing, and gives no Cartesia-only figure. Cartesia's homepage says it ranks first on an independent speech leaderboard. Neither is a substitute for listening to your own script in both.

Latency and real-time voice agents

Both vendors now target voice agents, and both publish speed numbers. They measure different things.

  • ElevenLabs quotes a median inference latency of about 100 ms for Eleven v4 Turbo and Flash v2.5, excluding application and network time. Its v4 post adds a separate time-to-first-speech figure measured over WebSocket.
  • Cartesia says Sonic produces first audio in under 90 ms.

Those numbers are close enough that your network path and your LLM will matter more than the TTS gap. Measure end to end, from the user stopping speaking to the first audio byte, in the region your callers are in.

ElevenAgents vs Managed Agents

ElevenAgents is the more complete platform today. It combines ElevenLabs' speech recognition, an LLM of your choice (or your own), its TTS and a proprietary turn-taking model. On top sit a visual workflow builder, a knowledge base, tool calling, SIP trunking, a native Twilio integration, batch outbound calls, client SDKs for web and mobile, and a hosted MCP server. ElevenLabs also offers speech-to-speech Realtime models for agents.

Cartesia's agent product is Managed Agents: Ink for transcription, an LLM you pick, Sonic for speech, with turn-taking, tools and telephony handled for you. You configure it in the Playground or through the API.

One status change to know if you're on Cartesia already. Cartesia's Line SDK migration guide says it stops hosting Line SDK agents on January 1, 2027. That covers code-first Python agents deployed to Cartesia. Self-hosted agent code keeps working, phone numbers carry over, and the guide says per-minute agent rates are unchanged. The Line SDK itself stays open source on GitHub. Many 2025 tutorials still tell you to deploy a Line agent to Cartesia's cloud, so check the date before you follow one.

If you want a phone agent without wiring the parts yourself, our Vapi vs Retell AI comparison covers platforms that can sit on top of either voice.

Speech-to-text on each side

Both vendors now sell transcription, which matters if you want one bill for the whole voice loop.

ElevenLabs Scribe v3 handles recorded audio, and Scribe v3 Live handles real-time audio. ElevenLabs' speech-to-text docs say Scribe v3 supports 90+ languages. Scribe v1 is deprecated. There's also Scribe v2 Medical for clinical audio.

Cartesia Ink 2 is a streaming model with built-in turn detection, aimed at voice agents. Cartesia's STT models page lists Ink 2 for English, French, Hindi, Japanese and Spanish. An ink-preview model released on November 17, 2026 covers 44 languages, but Cartesia says it can change without notice and recommends Ink 2 for production.

So for transcription outside those five languages today, ElevenLabs' Scribe is the safer production choice. If transcription is the main job rather than a side part of an agent, a dedicated STT vendor may fit better. Our Deepgram vs AssemblyAI post covers that comparison.

Pricing models compared

Both are usage-based, and both pricing pages change often, so here's the shape rather than a rate card.

ElevenLabs bills API text-to-speech per 1,000 characters, and the rate depends on the model. The ElevenLabs API pricing page puts Eleven v4 Turbo, v3 Conversational and Flash at roughly half the per-character rate of v4, v3 and Multilingual v2. Subscriptions run from a pay-as-you-go option through several monthly tiers to Enterprise. ElevenAgents is billed per minute plus LLM usage, and Scribe per hour of audio.

Cartesia bills text-to-speech in credits at one credit per character, which its pricing page says works out to roughly 750 to 800 credits per minute of audio. The Cartesia pricing page lists a no-cost entry plan, low-priced Pro and Startup tiers, Scale, and Enterprise. Voice agents draw from a separate prepaid balance billed per minute, with telephony and LLM usage on top. Instant voice replication starts on the first paid tier, and professional voice replication on Startup.

The practical takeaway:

  • For agents, compare the all-in per-minute cost on both, including the LLM and telephony.
  • For long-form content, compare v4 against Sonic on your actual character count, and check whether Flash or Turbo is good enough for your use. If it is, it roughly halves the ElevenLabs bill.

Which should you pick?

Your situationBetter fit
Audiobooks, narration, podcasts, dubbingElevenLabs
Expressive characters with directed deliveryElevenLabs, Eleven v4
Real-time phone or web agent, latency firstCartesia, Sonic 3.6
A language outside Sonic's 44ElevenLabs
Full agent platform with workflows, knowledge base and TwilioElevenLabs, ElevenAgents
Must run on-premise or in your own VPCCartesia, via Enterprise sales
Production audio must never change between releasesCartesia dated snapshots, or pin a specific ElevenLabs model ID

Where each one is honestly stronger:

  • ElevenLabs is stronger on language count, voice library, performance control, transcription coverage and the breadth of its agent platform.
  • Cartesia is stronger on a focused real-time story, model snapshots you can pin to a date, a documented on-premise option, and simple per-character billing.

If neither fits, we keep roundups of ElevenLabs alternatives and Cartesia alternatives, including open-weight models you can run yourself.

Where Taku Fits

ElevenLabs and Cartesia are building blocks. Someone still has to turn a voice API into a product you can use. If you'd rather run a tool someone has already built, that's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.

Key Points

  • ElevenLabs' current flagships are Eleven v4 and Eleven v4 Turbo, launched September 28, 2026, with 90+ languages. Turbo v2.5 is deprecated; Flash v2.5 is not.
  • Cartesia's current flagship is Sonic 3.6, with 44 languages and dated snapshots for production.
  • ElevenLabs quotes about 100 ms median inference for v4 Turbo; Cartesia claims first audio in under 90 ms. The figures are measured differently, and your network and LLM matter more than the gap.
  • ElevenAgents is the broader agent platform. Cartesia's Managed Agents replaces hosted Line SDK agents, which Cartesia stops hosting on January 1, 2027.
  • ElevenLabs says Scribe v3 supports 90+ languages. Cartesia's production Ink 2 model covers five, with a 44-language preview.
  • ElevenLabs bills per 1,000 characters by model; Cartesia bills one credit per character. Compare all-in costs for your workload.

FAQ

Is Cartesia better than ElevenLabs?

For low-latency voice agents, Cartesia is built around speed and offers dated model snapshots and an on-premise option. For expressive content, a wider language list and a larger voice library, ElevenLabs is stronger. Test both with your own script and network path.

Is Cartesia faster than ElevenLabs?

On paper they're close. Cartesia says Sonic produces first audio in under 90 ms. ElevenLabs quotes about 100 ms median inference for Eleven v4 Turbo and Flash v2.5. The two figures are measured differently, so run your own end-to-end test.

What is ElevenLabs' newest model?

As of November 2026, Eleven v4 for expressive content and Eleven v4 Turbo for agents and other low-latency use, both released on September 28, 2026.

What is Cartesia's newest model?

Sonic 3.6 is Cartesia's current TTS model. Its newest speech-to-text model in production is Ink 2, with an Ink preview model for 44 languages released on November 17, 2026.

Is ElevenLabs Turbo v2.5 deprecated?

Yes. ElevenLabs' models page marks Turbo v2.5 as deprecated and names Flash v2.5 as its replacement.

What happened to Cartesia's Line SDK?

The SDK is still open source, but Cartesia stops hosting Line SDK agents on January 1, 2027. Hosted agents should move to Managed Agents. Self-hosted agent code keeps working.

Can I replicate my own voice on both?

Yes. Both say an instant custom voice needs about 10 seconds of audio, and both offer professional voice replication for higher quality. On Cartesia, instant voice replication starts on paid plans.