Vendor-Claimed Text-to-Speech Latency 2026: Rime 37ms, Cartesia, ElevenLabs Turbo

Rime's Mist v3 has the lowest claimed time to first audio of the three providers compared here: approximately 37ms at the 50th percentile, at one concurrent request. Cartesia says its Sonic model delivers first audio in under 90ms, and ElevenLabs says its Eleven v4 Turbo has a median time to first speech of about 150ms. All three figures come from the vendors, and they do not measure exactly the same thing.

Which text-to-speech API has the lowest claimed latency in 2026?

Rime's Mist v3 has the lowest claimed figure. Rime's model documentation describes it as the model to use when you need the fastest time to first audio, citing approximately 37ms P50 in Rime's own benchmark. Rime's pricing page gives the same 37ms P50 at one concurrent request, with a P90 of 56ms, as of 13 October 2026.

Cartesia and ElevenLabs both report figures that are higher than Rime's, but the rows above are not a like-for-like race. Each vendor chose its own metric, percentile and test condition.

What does "time to first audio" actually measure?

Each vendor defines latency differently, so the numbers sit on different scales. Rime's 37ms is a P50 time to first audio at a single concurrent request. Cartesia's Sonic page gives first audio in under 90ms without a percentile or concurrency condition. ElevenLabs' Eleven v4 page reports a median inference latency of about 100ms and a median time to first speech of about 150ms.

The gap between model time and what a user hears is larger than any of these figures suggest. Rime says its cloud API delivers sub-200ms end-to-end latency under typical conditions when an application uses the nearest regional endpoint, according to its streaming documentation. End-to-end includes network time and application overhead. A voice agent builder should budget for that number, not the model-side figure alone.

Does Cartesia's under-90ms claim hold up against ElevenLabs' comparison?

The evidence available does not settle it. ElevenLabs' own v4 page compares its Eleven v4 Turbo at a 150ms median time to first speech with Cartesia Sonic 3.6 at 262ms and OpenAI GPT-4o mini TTS at 814ms. Cartesia's claim of under 90ms does not name a model version in the wording on its page, so the two figures may refer to different versions, different conditions, or different metrics.

Both companies are comparing their own products. ElevenLabs' comparison favours ElevenLabs, and Cartesia's headline favours Cartesia. Treat each figure as a claim until someone runs the same test on the same version.

How do the three APIs compare on languages and voices?

Rime is the narrowest of the three on language coverage, with four languages. ElevenLabs lists the broadest set. If your product needs a language outside these lists, check the vendor's current page before you commit, because each vendor updates its model lineup.

What are the production risks beyond latency?

Deepgram reports that ElevenLabs sets concurrency limits by tier: 2 concurrent requests on Free, 10 on Pro, and 15 on Business and Scale. Deepgram argues that these caps can force architectural compromises for applications with many simultaneous connections. The same Deepgram analysis says ElevenLabs models are non-deterministic, so identical inputs can produce different tone, volume and voice characteristics, which a stability control helps manage.

Cartesia's pricing page offers 20,000 free credits per month and lists paid plans from $5 per month for text to speech, as of 22 October 2026. CloudTalk lists a $4 per month Pro plan as of 21 May 2026, so the entry price is not settled. On voice cloning, AI Tool Lab reports that Cartesia's clones handle declarative speech well but flatten angry and excited tones. That is a single review, so weight it accordingly.

Which API should a voice agent team choose?

If latency is the only constraint and you need English, French, German or Spanish, Rime's claimed figure is the lowest in this comparison. If you need more languages, the choice moves to Cartesia or ElevenLabs. Haass reports that ElevenLabs' newest model, Eleven v4, leads Artificial Analysis' Provider Voice Arena ahead of Cartesia's Sonic 3.6, as of 6 October 2026. Voice quality and latency rank differently, so test on your own audio before you decide. For a closer look at ElevenLabs' models, read ElevenLabs v4 vs v4 Turbo: Is the #1 TTS Model Actually #1?

For Cartesia's leaderboard claims, see Cartesia Claims #1 on Artificial Analysis. The Leaderboard Says #4.

What's still unclear?

• Rime's 37ms is self-reported. It is measured at one concurrent request, and the vendor pages give no figure at higher concurrency. Nothing independent confirms it.
• Cartesia's version is unclear. The under-90ms claim does not name a model version, while ElevenLabs measured Sonic 3.6 at 262ms. Until Cartesia names the version behind its figure, the two numbers cannot be reconciled.
• Metrics differ. The three providers report time to first audio, time to first speech and end-to-end latency. A reader comparing the headline numbers is comparing definitions as much as speed.
• Cartesia's entry price is inconsistent across sources. The vendor's pricing page and a third-party listing give different starting prices for paid plans.
• The weakest point of this comparison is that it ranks vendor claims. A ranking of self-reported numbers is only as good as the methods behind them, and none of the three vendors publishes a method that a third party has reproduced.