On September 28, ElevenLabs released Eleven v4 — their new flagship voice model — alongside a faster v4 Turbo. More expression control, support for 90+ languages, and voice cloning from ten seconds of audio. Two weeks later they launched in Singapore, planting the flag three hundred kilometres from where I'm sitting.
I build systems for businesses that still run their front desk on a phone and a notepad. So I don't read voice AI announcements the way a tech blog does. I read them asking one question: can a clinic in Petaling Jaya afford this yet, and should they?
What Actually Shipped
The headline numbers from the official announcement: v4 Turbo answers with roughly 150 milliseconds to first speech, which is fast enough that a caller doesn't notice they're talking to software. Voice cloning needs ten seconds of sample audio. And TechCrunch's coverage confirms the 90+ language support — which matters here more than almost anywhere, and I'll get to why.
October kept rolling: an agent-building assistant called Architect, government-grade security certification, and a steady drip of agent platform features — call queueing, workflow outcome tracking, ticket merging. This is a company building a phone department, not a toy.
The Full Cost Stack, Honestly
The official pricing page reads like SaaS pricing always does: free tier to play with, $6 starter, $22 creator, $99 pro, up to $990 for business with per-minute rates quoted as low as five US cents. Roughly a thousand credits buys a minute of speech.
Here's what the pricing page undersells: text-to-speech is one line of a three-line bill. A working phone agent also pays for the language model doing the thinking and the telephony carrying the call. Third-party breakdowns put agent platform fees at around eight US cents per minute before those extra lines — I haven't verified that number against an official page, so treat it as indicative. The honest summary: a voice agent for a small business is now a hundred-dollars-a-month-class decision, not an enterprise project. That's new. It wasn't true a year ago.
The Southeast Asia Unlock
Ninety-plus languages is an abstract spec sheet line in California. In Malaysia it's the whole product. A front desk here answers in Bahasa, English, and Mandarin — sometimes in the same call, sometimes in the same sentence. One agent that handles that mix is the difference between a demo and a tool.
My caveat, from someone who will have to stand behind the invoice: verify the Malay quality yourself on the free tier before you promise it to anyone. Vendor language counts and real Manglish are different sports. I'd run twenty test calls before a single customer hears it.
The Dark Side of 10-Second Cloning
Ten seconds of audio is less than your last voice note. If your voice can be cloned from a WhatsApp message, then "the boss called and asked me to transfer the deposit" stops being proof of anything. This isn't theoretical — it's the obvious next move for scammers, and small businesses with informal payment approval are the softest target.
The fix costs nothing: a standing rule that unusual payment requests get verified by calling back on a known number. If I build you a system this year, that policy ships with it. The AI can talk — your process decides.
Where I'd Actually Use It
After-hours booking for clinics and salons. Missed-call capture for shops where the owner is the only phone-answerer. Order status lines for businesses drowning in "where's my parcel" calls. Boring, high-volume, low-stakes conversations — the exact work nobody enjoys and everybody pays a salary for.
What I wouldn't do: put it in front of anything involving money or medical advice, or deploy it without a one-tap path to a human. The technology got cheap. Judgement is still the expensive part.