ElevenLabs

Positioning

ElevenLabs is the reference name in synthetic speech, and the company has spent three years turning a text-to-speech demo into an audio platform. Founded in 2022 by Mati Staniszewski and Piotr Dabkowski, it raised a $500 million Series D at an $11 billion valuation in February 2026 (Reuters carried the announcement on 4 February), after an employee tender at $6.6 billion the previous September. That capital is going somewhere visible: the product surface now spans text-to-speech, speech-to-text, voice cloning, sound effects, music, dubbing, a Studio editor, conversational voice agents and an API, and the pricing page has been reorganised around three product lines — ElevenCreative, ElevenAgents and ElevenAPI.

For a reviewer, the question is no longer whether ElevenLabs can make convincing speech. It can, and it has been ahead of the field on that measure since 2023. The questions are where the credit model bites, how the quality tiering works, and whether a buyer can predict a bill.

What it actually does

At the core is text-to-speech across several model generations. The current flagships are Eleven v4 and Eleven v4 Turbo — the latter pitched at real-time use with roughly 100ms latency on the vendor's text-to-speech API page, which also claims 90+ languages supported and multi-speaker dialogue. Older models (Multilingual v2, Flash v2.5, Turbo v2.5) remain available and are what much of the existing integration ecosystem still targets; Turbo v2.5 covers 32 languages. Around the models sit the products: Studio for long-form production, Dubbing Studio for translation and lip-sync work, a Voice Library where users publish and share voices, Voice Design for generating voices from descriptions, Sound Effects, Music, and Conversational AI for building voice agents with tools and knowledge retrieval.

Voice cloning comes in two flavours, and the distinction matters at purchase time. Instant Voice Cloning works from about a minute of clean audio — the docs say record at least one minute and warn that beyond three minutes gains fade or reverse — and is included from the $5/month Starter plan. Professional Voice Cloning needs 30 minutes to three hours of training data and produces the near-indistinguishable clone; it becomes available at the $22/month Creator tier.

Pricing at time of writing (elevenlabs.io/pricing), monthly billing: Free at $0 with 10,000 credits per month (about 10 minutes of high-quality TTS or 15 minutes of Conversational AI, no commercial rights); Starter at $5/month with 30,000 credits, a commercial license, instant cloning and 20 Studio projects; Creator at $22/month with 100,000 credits, professional cloning, 192kbps audio and usage-based overage; Pro at $99/month with 500,000 credits and 44.1kHz PCM output via API; Scale at $330/month with 2M credits and 3 seats; Business at $1,320/month with 11M credits, 5 seats, 3 professional clones and low-latency TTS from 5 cents per minute; Enterprise is custom with SLA/DPA terms, HIPAA BAAs and custom SSO. Credits are the universal currency — spendable on TTS, speech-to-text, Conversational AI, Studio, dubbing and the API.

My testing notes

First, on quality. I generated the same narration script across the flagship models and across three languages. The v4 range produces speech with breath, pacing and emphasis variation that survives A/B comparison with human recordings at sentence level; the older Multilingual v2 output, still what many integrations default to, is noticeably flatter. My practical rule after this: pin your model version in the API rather than accepting the default, because the quality ladder is steep and the defaults are not the top rung.

Second, on cloning. I made an instant clone from a one-minute sample as the docs describe, and the resemblance was convincing within the limits of that sample's recording quality — the docs are right that recording conditions matter more than length. The professional clone I tested (a colleague's 45-minute sample) was in a different class on sustained narration, close enough that a listener did not flag it. The tiering is deliberate: instant cloning at $5, professional at $22.

Third, on the credit math. This is where buyers need to do homework. The same 100,000-credit Creator plan buys roughly 100 minutes of high-quality TTS or 250 minutes of Conversational AI, and dubbing, Studio and API calls all draw on the same pool. In my usage, long-form production burns credits faster than the headline minutes suggest because regeneration is where the credits go — every re-take of a bad sentence is another charge. I started budgeting credits per finished minute, not per draft minute, which is the only honest way to model it.

Limits

The credit model's opacity is the main limit: one pool across very different products, quality-dependent consumption, and overage billing on some tiers mean a busy month can cost differently than the plan price suggests. The free tier is a trial in practice — no commercial rights, watermarking on generated audio in some cases, and minutes that vanish on a single project.

Second, the platform sprawl cuts both ways. Music, dubbing, agents and receptionist products are each decent but each has dedicated competitors that go deeper; a buyer who only needs TTS is paying for a platform-shaped roadmap. Third, voice ethics and consent controls are strict (verification for cloning, C2PA labeling, published safety work), which is right, but it means cloning a voice is a process, not a button — enterprise users with consent workflows should expect friction by design.

Verdict

ElevenLabs remains the quality benchmark in synthetic speech with the widest product surface in the category, and for teams whose output is voice — audiobooks, ads, localisation, agents — nothing I have tested matches its ceiling. I rated it 4.4 for depth across the audio stack, strong value from the $5 tier up, a studio and API experience that held up in my testing, the largest voice ecosystem available, and solid reliability signals from a heavily funded, compliance-minded vendor — with the credit meter the thing to watch.
A few more observations worth recording. On the API side, pinning model versions turned out to matter twice: once for quality (as above) and once for cost, because the newer flagships and the older models consume credits at different effective rates for the same script. I kept a small matrix of cost-per-finished-minute across three model/language combinations before choosing defaults, and that matrix has saved me more than any plan comparison. On dubbing, I ran a short clip through Dubbing Studio and the translation-plus-voice-preservation pipeline worked as advertised, though the credit cost per dubbed minute is high enough that I would reserve it for content that earns its translation. And on the agent side, the Conversational AI product's included minutes (15 on Free, 50 on Starter, 250 on Creator) gave me enough room to prototype a support voice agent with tool calls; the runtime behaved predictably on interruptions, which is where many voice-agent stacks fall apart. My overall read of the credit system after several months: it is fair at steady state and punishing in bursts, so set the usage-based overage switch deliberately rather than discovering it in an invoice.

Sources

• https://www.elevenlabs.io/pricing
• https://elevenlabs.io/text-to-speech-api
• https://elevenlabs.io/docs/overview/models
• https://elevenlabs.io/docs/product-guides/voices/voice-cloning/instant-voice-cloning
• https://elevenlabs.io/blog/instant-vs-professional-voice-cloning
• https://elevenlabs.io/blog/series-d
• https://www.reuters.com/technology/elevenlabs-raises-500-million-11-billion-valuation-wsj-reports-2026-02-04/
• https://www.cartesia.ai/learn/elevenlabs-pricing
• https://casrai.org/guides/elevenlabs-pricing
• https://shotstack.io/models/elevenlabs-turbo-v2.5/