TTS API Pricing per Million Characters: The Arithmetic Behind the Rate Card

Verdict first

On the sticker price, Google Cloud TTS and Amazon Polly are the cheapest entries at $4 per million standard characters. On an API-first per-character basis, OpenAI and Deepgram sit at $15 per million for their entry models. Cartesia lands at $50 per million on Pro, or $37.375 on Scale, because the plan includes a credit pool. ElevenLabs is the premium option at roughly $165 to $200 per million, depending on the plan. The four vendors that sell minutes or credits without a character ratio cannot be placed in the same table without an assumption.

The per-million table

Every figure is derived from the vendor’s own price page. The formula column shows the arithmetic so a reader can check it.

The ElevenLabs and Cartesia rows assume one credit equals one character. Both vendors state that ratio for the models described on their pricing pages, so the conversion is not an outside guess. If you use a model with a different credit multiplier, the per-million figure changes.

Why the tiers are not interchangeable

A $4 Standard voice and a $30 Chirp 3 HD voice are not the same product. The cheap tier usually has a smaller voice set, a lower sample rate, or fewer controls. A per-million table is a map of the price ladder, not a claim that every rung sounds the same. The same applies to Neural, Generative, and Long-Form tiers: each is a different engine with a different price.

That is why the table lists the tier name next to the rate. A buyer who compares the $4 row with a competitor’s premium model is comparing a compact car with a sedan. The honest comparison is tier to tier, and even then the voice sets differ.

The first million is not the list price

Free allowances change the effective cost at low volume. Google gives 4 million Standard characters per month, plus 1 million Neural2 and 1 million Chirp 3 HD characters (pricing). Polly gives 5 million Standard characters per month (pricing). A one-million-character workload therefore costs nothing on those standard tiers.

OpenAI publishes no free allowance on its price page, so the first million costs $15 on tts-1 or $30 on tts-1-hd (pricing). Deepgram gives a $200 new-account credit, which at $30 per million covers several million Aura-2 characters before the first invoice (pricing). Cartesia gives 20,000 credits per month on Free and 100,000 on Pro; ElevenLabs gives 10,000 credits per month on Free.

The allowance matters most for a prototype or a low-volume product. At 100 million characters a month, the free tier is a rounding error and the list rate is the number that matters.

The credit overage trap

A credit plan has a second price: the overage rate. Cartesia publishes $65 per million credits on Pro when overage is enabled (pricing). That is higher than the $50 effective rate inside the plan, which is normal: the included pool is cheaper than the marginal credit.

ElevenLabs publishes plan credit pools but no overage rate on the pricing page reviewed. Once you exceed the top plan’s pool, the next step is a sales conversation rather than a published rate. That is a real difference in predictability, even if the headline per-million figure looks similar across plans.

The vendors without a character unit

Typecast sells 30,000 credits for $5 on Basic, described as about 35 minutes, and 3,000 one-time credits on Free, described as about five minutes (pricing). Without a characters-per-minute figure, a per-million price would require an assumption about speaking rate.

WellSaid sells download minutes: a three-minute monthly trial, then plans priced in dollars per month (pricing). Uberduck sells credits by plan: 1,000 on Starter, 3,600 on Creator, 25,000 on Pro (pricing). Neither publishes a character ratio on the page reviewed.

Respeecher publishes $2 per hour (product page). To estimate a per-million figure, assume a narration pace of 150 words per minute and 5.5 characters per word: 150 × 5.5 = 825 characters per minute, or 49,500 per hour. Then $2 ÷ 49,500 × 1,000,000 = $40.40 per million. That is an estimate built on two assumptions, not a vendor rate, and this site does not use it in the ranking.

Who should use which number

Prototypes should use the free allowance and ignore the per-million table until they exceed it. Google and Polly are effectively free at one million characters a month.

Products with predictable volume should use the per-character rate and model the monthly bill directly. OpenAI and Deepgram are the simplest to forecast because the rate is linear and the unit is characters.

Products on a credit plan should model both the included pool and the overage rate. Cartesia publishes both; ElevenLabs publishes the pool but not the overage rate on the page reviewed.

Products that need a specific voice should not start with the cheapest row. A $4 voice that does not fit the brand is more expensive than a $30 voice that does, once the rework is counted.

The takeaway

The per-million table is arithmetic, not a recommendation. Google and Polly are cheapest at the standard tier; OpenAI and Deepgram are the simplest API-first per-character rates; Cartesia sits in the middle; ElevenLabs is premium. The vendors that sell minutes or credits without a character ratio cannot be ranked on this axis until they publish the conversion.