ElevenLabs API Pricing: Cost, Free Tier and Rate Limits Broken Down

Open the ElevenLabs pricing page and you see six plans, a pile of credits and a few dollar figures per thousand characters. Open your first invoice and the numbers feel different. The gap comes from three things the page spreads across several tabs: how a model turns characters into credits, which plan your account sits on, and how many requests you can run at the same moment. This article puts all three in one place, with worked examples, so you can estimate a bill before you write a single line of integration code.

💡 Price check: every figure below was checked in early October 2026. ElevenLabs changes rates and runs promotions often, so treat each number as a snapshot and confirm it in your own dashboard before you commit budget.

What the API Actually Charges

Credits and Dollars Are Different

ElevenLabs measures usage in two currencies. Subscription plans hand you a monthly pool of credits, and every generated character spends some of them depending on the model. The usage-based API price list works in dollars per 1,000 characters instead. Both describe the same audio, but they do not convert one to one, which is why a quick mental calculation often lands far from the real number.

For text to speech the rule of thumb is simple. A multilingual model such as Multilingual v2 spends about 1 credit per character. The Flash and Turbo models spend between 0.5 and 1 credit per character, which is why they are the budget pick for chat bots and phone agents.

Rates by Model

Here is the usage-based price list for text to speech, per 1,000 characters.

The v4 discount ends a few days after this check, so budget on the list price, not the promotional one.

Picking between these rows is mostly a question of what the listener notices. A narrated video or an audiobook gets played once and judged on emotion, so the $0.08 models earn their price. A bot that answers "your order has shipped" a thousand times a day is judged on speed, and a $0.04 model that responds quickly beats a richer voice that makes the caller wait. Many teams run both: a premium model for content people choose to listen to, and a Flash model for everything automated.

Three Cost Examples

English averages roughly 6 characters per word once you count spaces and punctuation, so multiply words by 6 for a fast estimate.

• Audiobook chapter, 10,000 words. About 60,000 characters. At $0.08 per 1,000 that is $4.80. On a $0.04 model it drops to $2.40.
• Product video voiceover, 500 words. About 3,000 characters, so $0.24 at $0.08 or $0.12 at $0.04.
• Support bot, 50,000 replies of 300 characters. That is 15 million characters: $600 at $0.04 per 1,000, or $1,200 at $0.08.

💡 Quick formula: monthly cost = characters per month ÷ 1,000 × price per 1,000 characters. Run it twice, once for your lowest expected volume and once for your highest.

Plans, Credits and Monthly Cost

Subscriptions are the other way to pay. Each plan sells a fixed monthly pool of credits, plus extras such as a commercial license, team seats and higher concurrency.

Six Plans Compared

Annual billing works out to two months free: you pay for ten and use twelve. Unused credits roll over for up to two months, capped at three times your monthly allocation, as long as the subscription stays active.

Cost Per Credit by Plan

Divide the price by the credits and the plans stop looking alike.

Two things stand out. The cost per credit falls about 17 percent from Starter to Pro, then goes flat. Past Pro you are paying for concurrency, seats and support, not cheaper audio.

The second point matters more. At 1 credit per character these figures sit above the usage-based rates in the first table. The two price lists are billed differently, so price your expected volume against both before you pick one, and check in your dashboard which billing mode your account uses.

Three Sample Monthly Budgets

Here is how three realistic projects land when you price them both ways. The usage-based column uses the list rates from the first table.

Whether you can buy usage-based rates on your account depends on your billing setup, so read this table as a comparison, not a quote. The pattern still holds: small and medium projects fit a plan comfortably, while a high-volume bot is a sales conversation, not a checkout page.

The Free Tier in Practice

What 10,000 Credits Buys

The free plan gives you 10,000 credits every month for $0. On a model that spends 1 credit per character, that is 10,000 characters: around 1,700 words, or roughly 11 minutes of finished speech at a normal reading pace. On a Flash model at half a credit per character, the same pool stretches toward 20 minutes.

That is enough for three jobs:

• Picking a voice for a project
• Testing that your integration sends text and receives audio correctly
• Building a prototype to show a client or a teammate

Limits to Expect

• No commercial license. Audio you publish or sell needs a paid plan.
• Lowest concurrency. Two simultaneous requests on multilingual models and four on Flash and Turbo.
• No room for batch jobs. One 10,000-word chapter at 60,000 characters would eat six months of free credits.

Treat any of these three as your upgrade signal. The moment you need to publish the audio commercially, run more than two jobs at once, or generate more than about 10,000 characters a month, the free plan stops being a plan and becomes a trial. Starter at $6 is the smallest step up, and it triples your monthly credits while adding the commercial license.

💡 Tip: spend free credits on test scripts of two or three sentences, never on full chapters. Settle voice, speed and stability first, then run the full text once.

Rate Limits and Concurrency

Limits on this API are mostly about concurrency: how many requests are running at the same moment, not how many you send per minute. A fast model that finishes in a second lets one slot serve many requests. A slow, long generation holds its slot for the whole run.

Limits by Plan

These are the published limits at the time of writing. Enterprise accounts get elevated limits, and newer models can carry ceilings of their own, so check the dashboard.

Go over your limit and the API either queues the request, which adds a small delay, or returns an HTTP 429 error. Plan for both:

1. Retry with exponential backoff.
2. Add random jitter so retries do not arrive in a wave.
3. Cap in-flight calls on your side with a semaphore, set a little below your plan limit.

A minimal retry wrapper looks like this in Python:

import random
import time

def with_backoff(call, retries=5, base=0.5):
    for attempt in range(retries):
        try:
            return call()
        except RateLimitError:
            time.sleep(base * 2 ** attempt + random.random() * 0.25)
    raise RuntimeError("still rate limited after retries")

Copy Prompt

Swap RateLimitError for whatever exception your HTTP client raises on a 429. The wait grows from half a second to eight seconds, and the random fraction keeps parallel workers from retrying in lockstep.

Sizing Your Concurrency

Use this estimate:

concurrent requests ≈ requests per second × seconds per generation

An app that peaks at 8 requests per second with 1.5 seconds per generation needs about 12 concurrent slots. On Flash models that fits inside Pro, which allows 20.

A voice agent is different, because each live call can ask for speech at the same time. With 25 simultaneous callers at peak you need 25 slots, and on Flash models only Scale or Business (30) gives you that headroom. Plans are cheap compared with dropped calls, so size for your busiest hour, not your average one.

Transcription and Music Rates

Voice is only part of the bill. If your product also turns audio into text or makes soundtracks, those meters run too.

Scribe Speech to Text

Transcription is billed per hour of audio, not per character:

• Scribe v2: $0.22 per hour for files
• Scribe v2 Realtime: $0.39 per hour, with roughly 150 ms latency
• Vocabulary hints: $0.05 per hour extra
• Transcript editing: $0.07 per hour extra

A backlog of 100 hours of recorded interviews costs about $22. A two-hour live event transcribed in real time costs about $0.78. At those rates transcription is the cheapest line item in most audio products, so it rarely decides the plan. If you want a second opinion on accuracy, run the same file through GPT-4o Transcribe or Gemini 3 Pro and compare the transcripts.

Music and Voice Tools

Music generation is billed at $0.15 per minute of output. Thirty three-minute tracks is 90 minutes, or $13.50. Voice Isolator and Voice Changer each run $0.12 per minute.

Per-minute billing makes music easy to predict and easy to overspend, because every retry of a prompt is billed again. Write a tight brief, generate a short sample, and only then ask for the full length. You can audition ElevenLabs Music on PicassoIA before you wire it into code.

Ways to Cut Your Bill

Most savings come from sending fewer characters and picking cheaper models for the right jobs.

• Match the model to the job. Use Flash for bots and alerts, and save the premium models for narration where emotion matters. Halving the rate on bot traffic is usually the biggest single saving.
• Cache every repeated phrase. Greetings, menu prompts and error messages never change. Hash the text, voice and settings together, store the audio file, and serve it from storage. If 40 percent of your replies repeat, you stop paying for 40 percent of them.
• Send clean text. Strip HTML, markup, footnotes and repeated whitespace before the request. Every character is billed.
• Chunk long scripts with context. Split at paragraph breaks and pass the neighboring sentences as context, so tone stays consistent and a one-line edit does not force you to regenerate the whole script.
• Preview before you batch. Run one paragraph, listen, then run the rest.
• Set a ceiling in your own code. Track characters per day, alert at 80 percent of budget and stop at 100.
• Watch your retries. A retry loop that resends a request that actually succeeded bills you twice. Tag each job with an ID so you can spot duplicates.
• Mock the API in tests. A test suite that calls the real endpoint on every commit quietly burns credits. Record one sample response and replay it in continuous integration.
• Separate development from production. Staging environments, demos and experiments all draw from the same pool as your live product. Give them a tiny quota, or a dedicated account, so a loop in a test script cannot eat the month's allowance.
• Audit the biggest scripts monthly. Sort last month's generations by character count. A handful of long jobs usually account for most of the bill, and those are the ones worth caching or shortening.

Test Voices on PicassoIA First

Before you wire a paid API into production, audition voices where a wrong choice costs nothing. PicassoIA hosts several ElevenLabs models in its text-to-speech collection, plus models from other labs, so you can compare delivery side by side. It also offers ElevenLabs Dubbing for translating videos into other languages.

Models Worth Comparing

Step by Step with V3

1. Open the Eleven v3 page in the text-to-speech collection.
2. Paste two or three sentences of your script into the Prompt field.
3. Pick a voice. There are 26 to choose from, and the default is Rachel.
4. Set the language code, such as en, es or fr.
5. Leave stability at 0.5 and similarity boost at 0.75 for the first run.
6. Run it, listen, adjust one setting at a time, then download the audio.

Settings Worth Changing

💡 Tip: change one setting per run and keep notes. Ten short test clips teach you more about cost and quality than one long generation.

Try It on Picasso IA

Your next voiceover, soundtrack or transcript does not need a budget spreadsheet on day one. Open the text-to-speech collection on Picasso IA, paste a short script into V3, and hear the result in seconds. Then run the same script through two other voices and keep the one that fits your audience.

Building a podcast, a course or a product demo? The photorealistic image models on the platform can make the thumbnails, banners and scene art that go with your audio. Write a prompt, pick a model and see what comes out. Browse every voice, music, transcription and image model at picassoia.com/en/all-models, and start experimenting with your own projects today.