ElevenLabs just shipped a voice model built to shave the one thing that gives away every AI phone call: the before it answers.
On September 28, ElevenLabs released Eleven v4 and its faster sibling, Eleven v4 Turbo, pitching both squarely at the people building voice agents rather than the creators who made the company's name. Eleven v4 Turbo posts a median inference latency of about 100 milliseconds and a median time to first speech near 150 milliseconds over WebSocket streaming, according to ElevenLabs' own release. That's the gap between a voice bot that feels like it's thinking and one that just talks.
Here's the twist. Despite the "speech-to-speech" framing floating around this launch, Eleven v4 Turbo is not an integrated speech-to-speech model at all. ElevenLabs still runs its agents on a modular pipeline: speech-to-text, then a large language model, then text-to-speech, stitched together inside its ElevenAgents platform. OpenAI's Realtime API, by contrast, processes voice in and voice out through a single integrated model, no transcription step in between. ElevenLabs has written about this difference itself, in a blog post comparing its Conversational AI directly against OpenAI's Realtime API.
So the real news isn't a new architecture. It's that ElevenLabs has pushed its old architecture fast enough to compete with the new one on feel, while keeping what the single-model approach doesn't offer: an actual product. OpenAI's Realtime API gives developers raw low-latency voice infrastructure and stops there. No agent builder, no telephony hookup, no knowledge base, no built-in call analytics. ElevenLabs packages all of that inside ElevenAgents, with over 3,000 voices against OpenAI Realtime's six, plus tool integration, WhatsApp support, and content guardrails already wired in.
Eleven v4 Turbo isn't actually ElevenLabs' fastest model. Flash, the company's older low-latency option, streams first audio in roughly 75 milliseconds, against v4 Turbo's 150. What v4 Turbo adds instead is emotional range: it reads pacing, tone, and context well enough to get quieter when a caller sounds frustrated or speak more clearly when they sound confused, according to ElevenLabs. That's the trade the company is betting enterprises will make, accepting double the delay of Flash in exchange for a voice that doesn't sound flat on a twenty-minute support call. It also now covers more than 90 languages, with ElevenLabs citing its biggest gains in Japanese, Brazilian Portuguese, Mandarin, and Cantonese.
OpenCode breach let hackers burn users' AI credits for model theft OpenCode says a legacy Console system exposed stored OpenAI, Anthropic and Google keys for 1.4% of users, with individual accounts hit for sums from $281 up past $1,700 before providers revoked the keys. - how to protect stolen API keys from hackers - OpenCode breach drained users AI credits for theft
The pricing underneath hasn't moved. ElevenLabs bills $0.05 per 1,000 characters on its fast-tier models, Flash and now Turbo, against $0.10 on the higher-quality Multilingual v2 and Eleven v3 models. For a company burning through support call volume, that's a real line item, not a rounding error.
The money behind the bet #
ElevenLabs isn't building this from a position of weakness. The company closed a $500 million Series D in February led by Sequoia Capital, valuing it at $11 billion, more than triple its $3.3 billion mark from the November 2024 Series C. A third close in May pulled the round past $550 million and added BlackRock, Wellington, D.E. Shaw, Nvidia, Salesforce, and Deutsche Telekom as investors. More recently, the company has reportedly been in early talks for a secondary tender that would value it near $22 billion, double the February number, though that figure hasn't been confirmed by a signed deal.
That capital is the reason ElevenLabs can afford to specialize instead of bundle. Google and OpenAI both treat voice as one feature inside a much bigger model lineup. ElevenLabs has nothing else to sell. Every dollar of that $11 billion valuation rides on voice agents actually working at enterprise scale, which is exactly why a 50-millisecond latency improvement gets a product launch and a blog post instead of a changelog entry.
Frankly, the fight here isn't about who has the smartest model. It's about who owns the call center. Enterprises don't buy architecture purity, they buy whether the bot can handle an angry customer without sounding like a kiosk. ElevenLabs is betting that a modular pipeline tuned hard enough beats an elegant integrated model with no product around it, and for now, the pricing and the packaging say that bet is working.
Also read: Jensen Huang Warns Europe Not to Overthink AI Rules or Risk Falling Behind • Why 4Humanity Built a Searchable Record of the Good Already Happening Across America • Salesforce will buy Braze for $5.7 billion to feed its AI agents customer data
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.