Text-to-speech API pricing, compared per character

Text-to-speech vendors do not agree on a billing unit, which is the main reason their price lists are hard to compare. The most common unit is the character: you send text, the vendor counts the characters in it, and you pay a rate per 1M characters. Amazon Polly, Google Cloud, Microsoft Azure, OpenAI's tts-1 models, ElevenLabs and Deepgram Aura all work this way, although the rates differ by a wide margin.

A second group bills the audio that comes out instead of the text that goes in. OpenAI's gpt-4o-mini-tts is published as text input tokens plus audio output tokens, with a rule of thumb of roughly one and a half cents per minute of generated speech. Google's Gemini speech models are also token-priced: a rate per 1M text tokens for the input and a higher rate per 1M audio tokens for the output, at 25 audio tokens per second of speech. A few vendors quote a flat per-minute price.

Some vendors add their own twist. ElevenLabs sells characters as credits inside subscription plans and quotes a separate API rate. Fish Audio bills per UTF-8 byte, so English text counts about one unit per character while Chinese, Japanese or Arabic text counts several. Google Cloud splits its catalogue into voice tiers (Standard, WaveNet, Neural2, Chirp 3 HD and Studio) whose prices span a factor of forty, and Amazon Polly does the same across standard, neural, generative and long-form engines. In this table the 35 models come from 15 providers, and 6 of them are priced in tokens or minutes and therefore converted rather than read off directly.

Every row is converted to dollars per 1M characters of input text, the unit most vendors already use. Conversion needs an assumption about how much text makes a minute of audio, and this page uses 900 characters per minute: 150 words a minute, a normal conversational narration pace, times 6 characters per word including the space.

• Per-character prices are used as published, with no conversion.
• Per-minute prices become dollars per 1M characters as price per minute divided by 900, times 1,000,000. That is how gpt-4o-mini-tts is placed on the table.
• Token prices use 1,500 audio tokens per minute (25 per second) for the output plus 225 text tokens per minute for the input, which is 900 characters at 4 characters per token. The result is divided by 900 and scaled to 1M characters.

Rows converted this way carry a tilde in the table so they are not mistaken for a published rate. They are estimates, accurate only as far as the speech you generate matches 900 characters per minute. A slow, pause-heavy read costs less per character than the table suggests for token-priced models, and a fast read costs more. Right now 6 of 35 rows are estimates.

Take a 10-minute narration of about 9,000 characters. At hexgrad Kokoro-82M that costs $0.0089; at Google Cloud Studio it costs $1.44. Both are small numbers for a one-off job, which is why price differences only matter once volume arrives.

A support line that speaks to 1,000 callers for 2 minutes each generates about 1.8 million characters a month. That is $1.78 a month on hexgrad Kokoro-82M, $36 on the median model (hexgrad Kokoro-82M (US English)) and $288 on Google Cloud Studio. The priciest row costs about 162 times the cheapest.

These figures are list prices at the volume you type in. Markup counts too: if a vendor bills SSML tags or other markup as characters, the billed total is higher than the plain word count suggests. Retried generations, voice previews and re-recordings are billed like any other request, so budget for more characters than the final script contains.

The monthly figures use list prices, with no free tiers, volume discounts, committed-use contracts or prepaid credits. Free allowances can be large: Google Cloud gives away 4M characters a month on Standard and WaveNet voices and Amazon Polly gives 5M on its standard engine, and the table shows them in its own column so you can subtract them. Voice cloning fees, concurrency limits and latency are not part of the price either.

Quality is not scored here. A cheap voice is only cheap if you can ship it, and the only honest test is to listen to your own text. For blind-vote listening comparisons across vendors, see the Hugging Face TTS Arena (opens in new tab).

Open-weight voices are sold by several hosts at very different rates. The cheapest open-weight row here is hexgrad Kokoro-82M at $0.99 per 1M characters and the priciest is Canopy Labs Orpheus v1 English at $22. Because anyone can host the weights, the same model family often appears at several different prices, so compare hosts, not just models.

Dated changes matter. Google Gemini 3.8 Flash TTS is $15.13 today and rises to $30.25 per 1M characters on January 1, 2027, so a budget that covers next year should use the later number.

4 rows are marked as retiring, the earliest on Nov 17. Preview models tend to be withdrawn on short notice, so avoid building a long-term cost plan on them.

Prices come from vendor pricing pages and from the LiteLLM community price list, refreshed daily. 26 of 35 rows are marked Official, meaning the price was read from the vendor's own page and checked against it; the rest are labeled LiteLLM, Reseller or Third-party in the source type column, and the Verified only filter hides third-party and reseller prices. Each row links to its source and shows the unit the vendor actually publishes, so you can verify the conversion yourself. For the full process, see the methodology page.

Not listed: Hume Octave (no price could be verified), Resemble (published prices conflict between sources), PlayHT (service reported as discontinued), LMNT (service reported as shut down) and Replicate Kokoro (billed per run, not per character). Rows return when a verifiable price is available. Prices were last checked on the date shown at the top of the page.