TTS API pricing in 2026: how ElevenLabs, Deepgram, Cartesia, OpenAI and Google bill for speech

A TTS price card looks simple: a number per million characters. Then you put a voice agent into production and find out what the card left out. One vendor bills credits that roll over, another bills audio tokens, a third bundles characters into a monthly plan, and most cap how many requests you can run at once based on which plan you're on.

So this guide doesn't rank vendors by the lowest number. It explains how each of the five most-asked-about TTS APIs meters usage, what the concurrency caps are per plan, and what one minute of agent speech costs under each model. Every price below was pulled from the vendor's own pricing page and is dated, as of November 2026. Prices change, so check the linked page before you sign anything.

TTS API pricing at a glance (as of Nov 2026)

Sources are linked in each vendor section below. All prices are USD list prices, as of Nov 2026.

Four ways TTS APIs bill you

Before comparing numbers, it helps to know which of four meters you're on. They behave very differently once traffic grows.

1. Per character, pay as you go. You pay for the text you send, at a fixed rate. Deepgram, OpenAI's tts-1 models, Google's Cloud TTS voices and AssemblyAI's Text-to-Speech API all work this way. It's the easiest to forecast: count the characters your agent says, multiply. The fine print is what counts as a character. Google, for example, counts spaces, newlines and every SSML tag except <mark>.

2. Credits inside a plan. Cartesia sells monthly plans that include a block of credits, where one credit is one character of Sonic speech. Unused credits roll over up to 2x your monthly allowance. Go past your balance and you either pay a per-1M overage rate or, with overages off, new requests are blocked.

3. Monthly plans with included characters. ElevenLabs' API bills per 1,000 characters, but its subscription plans (Starter $6, Creator $22, Pro $99, Scale $299, Business $990 a month) include a character allowance that varies by model. The plan you pick also sets your concurrency limit, so the plan decision is a capacity decision, not just a discount.

4. Tokens. OpenAI's gpt-4o-mini-tts and Google's Gemini TTS models bill text input tokens plus audio output tokens. The audio side dominates, and it scales with how long the speech is, not how many characters you sent. That makes a per-minute estimate harder: the same sentence spoken slowly costs more.

Vendor by vendor

ElevenLabs

As of Nov 2026, ElevenLabs lists $0.08 per 1,000 characters ($80 per 1M) for Eleven v4, v3 and Multilingual v2, and $0.04 per 1,000 ($40 per 1M) for v4 Turbo, v3 Conversational and Flash/Turbo (elevenlabs.io/pricing/api). API usage is billed in dollars, not credits.

The limit to plan around is concurrency. ElevenLabs' docs set a per-plan cap on simultaneous requests: 2 on Free, 3 on Starter, 5 on Creator, 10 on Pro and 15 on Scale and Business for Multilingual v2, double that for Flash. Requests over the cap wait in a queue (elevenlabs.io/docs/overview/models). If you need more simultaneous speech, you move up a plan or go to Enterprise.

Deepgram

Deepgram bills per 1,000 characters of input text. As of Nov 2026, Pay as you go is $0.045 for Flux TTS, $0.030 for Aura-2 and $0.015 for Aura-1. The Growth plan, which starts at $4,000 a year in prepaid credits, lowers Aura-2 to $0.027 and Aura-1 to $0.0135. Flux TTS is the same on both (deepgram.com/pricing).

On concurrency, the pricing page says limits "apply per project and differ by plan and region" and publishes no TTS figures. Growth overages bill at the Growth rate plus 10%. Flux TTS carries a $1-for-$1 credit match up to $500 through Dec 31, 2026, so its effective price changes after that date.

Cartesia

Cartesia sells credits in plans. As of Nov 2026: Free includes 20K credits a month, Pro is $5 for 100K, Startup $49 for 1.25M and Scale $299 for 8M. One credit is one character of Sonic text-to-speech, which Cartesia puts at 750 to 800 credits per minute of audio. Overage is $65 per 1M on Pro, $45 on Startup and $38 on Scale (cartesia.ai/pricing).

TTS concurrency is fixed per plan: 2 concurrent requests on Free, 3 on Pro, 5 on Startup and 15 on Scale, custom on Enterprise. Break tags count as 1 credit each.

OpenAI

As of Nov 2026, OpenAI lists tts-1 at $15 per 1M characters and tts-1-hd at $30 per 1M. gpt-4o-mini-tts bills $0.60 per 1M text input tokens and $12 per 1M audio output tokens (developers.openai.com/api/docs/pricing).

The bigger issue is lifespan. OpenAI's deprecations page lists tts-1, tts-1-hd and gpt-4o-mini-tts for shutdown on January 6, 2027, with gpt-realtime-2.1-mini as the recommended replacement (developers.openai.com/api/docs/deprecations). That replacement is a token-billed realtime model, so a migration changes your meter, not just your model name. If you're on those models today, our guide to migrating off OpenAI TTS walks through the swap. Throughput is governed by OpenAI's account-level rate limits by usage tier.

Google Cloud Text-to-Speech

Google prices per character for its Cloud TTS voices and per token for Gemini TTS. As of Nov 2026: Standard and WaveNet voices are $4 per 1M characters with 4M free a month, Neural2 is $16 per 1M with 1M free, Chirp 3: HD is $30 per 1M with 1M free, Studio is $160 per 1M, and instant custom voice is $60 per 1M (cloud.google.com/text-to-speech/pricing).

Gemini 3.8 Flash TTS, in preview, is $0.50 per 1M text input tokens and $9 per 1M audio output tokens through December 31, 2026. On January 1, 2027 the listed price doubles to $1 and $18. Throughput is set by per-project Cloud TTS quotas.

AssemblyAI

AssemblyAI-TTS-1 is $15 per 1M input characters on pay as you go, as of Nov 2026 (assemblyai.com/pricing). Cloned voices and EU processing are the same price, and the voice library, instant voice cloning and word-level timestamps are included rather than sold as add-ons. One detail worth knowing: the character count includes text you sent that was passed to synthesis and then cancelled on barge-in, so trimming what you send matters more than trimming what gets played.

Concurrency: the cost that isn't on the price card

Price per character tells you what a minute of speech costs. Concurrency tells you how many minutes you can produce at the same moment. For a voice agent, the second number often decides the plan.

Here's why. Every agent turn is a TTS request. At 200 simultaneous calls, you might have dozens of replies being synthesized at any instant. On a plan capped at 5 or 15 concurrent requests, the rest wait in a queue or get rejected, and the caller hears that wait as dead air.

The caps fall into three shapes:

• Fixed per plan. Cartesia (2 to 15) and ElevenLabs (2 to 15 for Multilingual v2, 4 to 30 for Flash) publish hard numbers. Raising the cap means a bigger plan.
• Unpublished or account-tier. Deepgram says its limits differ by plan and region. OpenAI and Google govern throughput through account rate limits and project quotas. You find your real ceiling in a load test or a sales call.
• Session starts per minute. AssemblyAI counts new TTS sessions per minute, not sessions open at once. A voice agent opens one session per call and keeps it for the whole call, so calls in progress don't count against the limit. The starting limit is 5 new sessions a minute on a free plan and 200 on a paid plan, and on a paid plan it grows with use (TTS limits). The pricing page puts it plainly: "No spend commitment to unlock the concurrency you need."

Whatever vendor you pick, test it at your real peak before launch. Our guide to voice agent load testing covers how to run that test.

Pricing a voice-agent minute

To compare meters, convert everything to one unit: a minute of agent speech. We'll use 800 characters per minute, the top of Cartesia's own 750 to 800 estimate. ElevenLabs' calculator assumes closer to 1,000, so your real number depends on your agent's speaking rate. Measure it from your own transcripts.

Token-billed models (gpt-4o-mini-tts, Gemini TTS) are left out of the per-minute table because their cost depends on audio tokens per second of speech, which neither vendor lists next to the price.

Now scale it. Say your agent handles 10,000 calls a month and speaks for two minutes on each. That's 20,000 minutes of agent speech, or about 16M characters. At list price, that's $240 a month on AssemblyAI-TTS-1, $480 on Deepgram Aura-2, $640 on ElevenLabs Flash, and about $603 on Cartesia's Scale plan (8M included, then 8M at the $38 overage rate). Google Standard voices come out lower than any of them, at about $48 after the free tier.

TTS is also only one layer. A full call adds streaming speech-to-text and an LLM, so the TTS line is one of three. For the full per-call math, see our breakdown of AI voice agent pricing. On AssemblyAI the other two layers are Universal-3.6 Pro Realtime at $0.45 an hour, and any model through the LLM Gateway at provider rates with 0% markup.

What the price card won't tell you

A low-cost minute of speech that reads an order number wrong costs you a repeat turn, a frustrated caller and sometimes a human transfer. So once you've narrowed the list on billing model and concurrency, compare what you're actually buying:

• Entity accuracy. Does the voice read phone numbers, prices, dates and emails correctly?
• Time to first audio. How long does the caller wait before the agent starts talking?
• Price stability. Is the model in preview, on a promo, or scheduled for shutdown?

On AssemblyAI's own evals, published in the AssemblyAI-TTS-1 launch post, AssemblyAI-TTS-1 scored 93.7% entity accuracy on a pronunciation set, against 93.6% for ElevenLabs v4, 82.5% for Cartesia Sonic 3.6 and 68.0% for ElevenLabs Flash v2.5. The same evals measured a 47 ms median time to first byte. That's a lab number with a single client sending one request at a time; production time to first audio is about 90 ms.

Independent testing points the same way. On Cekura's TTS benchmark, which covers 104 agent sentences across 27 models, AssemblyAI-TTS-1 ranks #1 on phone numbers and codes at 91.7% and sits in the "Best Quadrant" for accuracy against latency.

If you're still choosing on quality, our roundup of the best TTS APIs for voice agents and the broader list of top text-to-speech APIs go model by model.

How to choose a TTS pricing model

Pick the meter first, then the vendor:

1. Measure characters per call. Pull a week of agent replies and count. Everything else multiplies off this number.
2. Find your peak concurrency. Look at your busiest minute, not your average. If a vendor's plan caps simultaneous requests below that, the plan price is not your real price.
3. Check the calendar. Promo credits, preview pricing and shutdown dates all change your cost in 2027. Three of the five vendors here have a dated change on the books: a promo ending, a preview price rising and a shutdown.
4. Price the whole call. TTS sits next to speech-to-text and the LLM. A low TTS rate on a stack that mishears callers isn't a saving.

Final words

TTS pricing in 2026 isn't one number. It's a billing unit, a plan, a concurrency cap and a date. Characters are the easiest to forecast, credits and plans tie capacity to spend, and tokens tie cost to how long the speech runs. Put all four on one sheet before you compare list prices, and the right choice for your traffic usually becomes obvious.

Frequently Asked Questions (FAQs)

How much does the ElevenLabs API cost per character?

As of Nov 2026, ElevenLabs lists $0.08 per 1,000 characters for Eleven v4, v3 and Multilingual v2, and $0.04 per 1,000 for v4 Turbo, v3 Conversational and Flash/Turbo. That works out to $80 and $40 per 1M characters. Subscription plans include a character allowance and set your concurrency limit.

How is Cartesia TTS priced?

Cartesia bills credits, where one credit equals one character of Sonic text-to-speech. As of Nov 2026, Pro is $5 a month for 100K credits, Startup $49 for 1.25M and Scale $299 for 8M, with overage at $65, $45 and $38 per 1M. Each plan also fixes TTS concurrency, from 3 requests on Pro to 15 on Scale.

How much does OpenAI TTS cost?

As of Nov 2026, OpenAI lists tts-1 at $15 per 1M characters, tts-1-hd at $30 per 1M, and gpt-4o-mini-tts at $0.60 per 1M text tokens plus $12 per 1M audio tokens. OpenAI's deprecations page lists all three for shutdown on January 6, 2027.

Is Google text-to-speech free?

Google Cloud Text-to-Speech has a monthly free tier, not a free product. As of Nov 2026, the first 4M characters of Standard or WaveNet voices and the first 1M of Neural2 or Chirp 3: HD voices are free each month, then billing starts at $4, $16 or $30 per 1M characters. Gemini TTS models have no free tier on the Cloud pricing page.

Does TTS API pricing include concurrency?

Usually not directly, but your plan often sets it. ElevenLabs and Cartesia publish per-plan caps on simultaneous TTS requests, and Deepgram says its limits differ by plan and region. AssemblyAI limits new TTS sessions per minute rather than sessions open at once, so calls already in progress don't count against the limit.

How many characters is one minute of speech?

Roughly 750 to 1,000 characters, depending on speaking rate and how much punctuation and number formatting the text contains. Cartesia puts it at 750 to 800 per minute, and ElevenLabs' calculator assumes about 1,000. Measure your own agent's replies for an exact figure.