How much does AI text to speech actually cost, and how do I pick a model?

This page may contain affiliate links from which Blur Studio earns commissions.

ElevenLabs sells several different speech models side by side, and the honest answer to "which one should I use" is that most people should default to the cheapest fast model and only pay more when a specific piece of audio is actually public-facing and benefits from the extra expressiveness. Picking the most expensive model as the default is the single easiest way to overspend on this.

The current models, compared

As of 29 October 2026, checked directly against ElevenLabs' own model documentation:

Model	Languages	Character limit per request	Latency	API price
Eleven Flash v2.5	32	40,000	~75ms	$0.04 / 1,000 chars
Eleven v4 Turbo	90+	Not separately published	~100ms	$0.04 / 1,000 chars
Eleven Multilingual v2	29	10,000	Standard (not optimized for speed)	$0.08 / 1,000 chars
Eleven v3	70+	5,000	Standard	$0.08 / 1,000 chars
Eleven v4	90+	10,000	Standard	$0.08 / 1,000 chars

The pattern is simple once you see it laid out: the fast models (Flash v2.5 and v4 Turbo) cost half as much per character and respond in well under a quarter-second, which matters for anything live, a voice agent, a real-time conversation. The standard-speed models cost double but generally handle more expressive delivery and longer-form narration more reliably.

A worked example: putting real numbers against this

Numbers make this concrete. Say you're generating roughly 20 minutes of finished audio a week across training content, product narration and the odd ad. Spoken English runs at roughly 800 to 1,000 characters per minute of audio at a natural pace, so 20 minutes a week is around 18,000 characters, or about 72,000 characters a month. On Flash v2.5 at $0.04 per 1,000 characters, that's roughly $2.90 a month in raw generation cost. On a standard model at $0.08 per 1,000 characters, it's about $5.80. At this volume, the difference between the cheap and expensive model is a few dollars, not a meaningful line item, which is exactly why I'd default to picking the model based on how public-facing the audio is, not on saving $3 a month.

The bigger number is usually the subscription tier, not the per-character model choice. The Free plan's 10,000 monthly credits covers roughly 10,000 characters, enough for light, occasional use but not for a business generating audio weekly. Starter's 30,000 credits a month covers that 72,000-character example only if you're on the cheaper Flash-class models; on standard models you'd burn through the Starter allowance in under two weeks and need to either upgrade or pay API overage rates. Work out your actual monthly character count first, then size the plan to it, rather than guessing and discovering the gap when you hit a wall mid-month.

Subscription credits vs. pay-as-you-go API billing

There are genuinely two different ways to pay here, and picking the wrong one for your usage pattern costs real money. A monthly subscription plan bundles a fixed number of credits for a flat fee, which is the better deal if your usage is steady and predictable, you're generating roughly the same volume every month, so the bundled rate beats paying per character every time. Pure pay-as-you-go API billing at the per-character rates in the table above makes more sense for spiky, unpredictable usage, a business that generates almost nothing most weeks but has an occasional large batch job, since you're not paying for a monthly credit allowance that mostly goes unused. If you're not sure which pattern describes you, check your actual usage over the last two or three months before committing to a plan tier; a one-off big project doesn't justify upgrading your standing monthly subscription.

What changed recently, and why it matters right now

Two things happened in the last few months that make some older comparisons of ElevenLabs' models out of date. Eleven v4 launched on 28 September 2026, expanding language coverage from the prior 70+ languages to over 90, adding voice cloning from as little as 10 seconds of reference audio, and introducing stackable inline expression tags carried over from v3. It was only added to ElevenLabs' public plan-comparison pricing page between 19 and 26 October 2026, so if you're reading a "best ElevenLabs model" article written before late October, it likely doesn't mention v4 at all, or buries it as a beta feature rather than a fully priced, generally available model.

Separately, on 12 July 2026, ElevenLabs deprecated its older Turbo v2.5 model in favor of Flash v2.5, which it describes as matching Turbo's voice quality at lower latency. If you're referencing an older integration guide or a platform's documentation that still names "Turbo v2.5" as the recommended fast model, that guidance is stale; Flash v2.5 is the one to use now, and existing Turbo-based agents on most platforms were automatically migrated.

How to actually pick one

1. Work out your real character count first. Pull a representative sample of what you're actually generating (a week's worth of audio, say) and count the characters, rather than guessing. This number, not the per-character rate, is what actually determines your bill.
2. Default to Flash v2.5 for anything that doesn't need maximum expressiveness: internal content, draft narration, high-volume or real-time use. It's half the price and the quality gap versus the standard models is small for most everyday content. Speed matters most on calls, as in why your AI phone agent replies slowly.
3. Reserve the standard models (v4, Multilingual v2, v3) for public-facing, brand-sensitive audio where the extra expressiveness and longer per-request character limit genuinely matter, a hero ad, a customer-facing video, long-form narration that needs to stay stable over several minutes. Long scripts can drift, as explained in why AI voice changes tone in long audio.
4. Test both on the same script before deciding rather than assuming the price difference implies a quality difference you'll actually notice. For a lot of straightforward narration, most listeners can't reliably tell the fast model apart from the expensive one.

How this compares to Google Cloud and Amazon Polly

Google Cloud Text to Speech and Amazon Polly both price per character too, and for their most basic, robotic-sounding voice tiers they're generally cheaper than ElevenLabs. Where that comparison breaks down is voice quality and expressiveness: ElevenLabs' models are built specifically around sounding natural and emotionally expressive, which is the entire reason most people reach for it over a cloud provider's default TTS in the first place. If raw cost per character is your only constraint and voice quality genuinely doesn't matter for your use case, it's worth pricing both directly on their own current pages before committing, since cloud-provider pricing pages change independently of ElevenLabs'. A closer look is ElevenLabs vs Amazon Polly vs Google text to speech.

Worth knowing if you're shopping around: October 2026 saw a run of developer discussion directly comparing ElevenLabs against newer, often cheaper voice models like Fish Audio, Voxtral and Google's Gemini TTS. None of these beat ElevenLabs outright on every dimension, but the gap in raw voice quality that used to make ElevenLabs an easy default has narrowed enough that it's worth a side-by-side test on your actual script before assuming ElevenLabs is the only serious option. Another comparison is ElevenLabs vs OpenAI text to speech.

What I'd actually do with that discourse, rather than chase every new model that shows up in a comparison thread: generate the same 30-second script on ElevenLabs' Flash v2.5 and on whichever newer competitor is getting attention, and listen to both back to back on a normal speaker, not studio headphones, since that's how most of your actual audience will hear it. If you can't reliably tell them apart for your specific content, the cheaper option wins by default. If ElevenLabs' version is noticeably more natural for your particular script, the extra cost is usually trivial next to the better result, and switching providers has its own real cost (re-recording and re-testing everything you've already built), which is worth weighing against a marginal price difference.

My honest take on model choice

Most people overthink this decision. The actual cost difference between Flash and the standard models, at the volumes most individual businesses generate, is a few dollars a month, not a budget-defining choice. Where I'd spend actual time deciding is the subscription tier, since jumping from Starter to Creator to Pro is a much bigger jump in both cost and included credits than switching models within the same tier. Pick the model based on whether the audio is public-facing (standard models) or internal/high-volume (Flash), and pick the subscription tier based on your actual monthly character volume, measured, not estimated.

When this is worth a custom build

If you're routing production traffic through ElevenLabs programmatically, picking the model dynamically based on context (fast model for a live call, standard model for a pre-recorded ad from the same system), that's a case for building a small routing layer rather than manually choosing a model each time in the dashboard. I build exactly that kind of connective logic for clients who've outgrown doing this by hand.