ElevenLabs ships v4 and v4 Turbo: 90+ languages, ~100ms Turbo latency

ElevenLabs released its v4 and v4 Turbo speech models on September 28, available immediately through the apps and API — not a preview. v4 raises language coverage from 70 to 90+, clones a voice from about 10 seconds of audio, restores Professional Voice Clones (missing in v3), and introduces stackable inline expression tags — [laughs], [whispers] and [pause] execute in sequence. v4 Turbo targets voice agents at a claimed ~100 ms median inference latency.

The facts

• v4: 90+ languages (v3 had 70); voice cloning from ~10 seconds of audio; Professional Voice Clones restored; stackable inline expression tags executed in order.
• v4 Turbo: built for voice agents. ElevenLabs’ own page compares ~100 ms median inference and ~150 ms time-to-first-speech against Cartesia Sonic 3.6 (262 ms) and OpenAI GPT-4o mini TTS (814 ms) — vendor-measured numbers.
• Availability: live today in the apps, the API (model name eleven_v4) and ElevenAgents; the free tier includes 10,000 credits per month (~10 minutes of audio), paid plans from $6/month.
• Context: both models were teased at a Warsaw event earlier in 2026. TechCrunch also cites funding reports ($500M at $11B earlier this year) and a revenue run rate growing from $330M to over $600M — press-reported, not disclosed filings.

Our take

The interesting part is less audio quality than latency and engineering: inline expression tags make “laugh here, pause there” versionable markup, and Turbo’s ~100 ms tier is aimed straight at conversational agents. Teams doing dubbing, audiobooks or voice bots should benchmark it directly — against local open-source tools like YoVoice it saves deployment at the cost of per-usage billing. One caution: the latency comparisons are the vendor’s own; run your own scripts through it before committing.