Per-Character Pricing: Examples & Companies

What is it

Per-Character Pricing is a billing unit where customers are charged per character of text processed — the standard meter for text-to-speech and translation. The vendor bills the text going in, not the audio or translation coming out: a 10,000-character script costs the same whether the voice reads it in six minutes or eight, which makes the unit fully countable before a single second of audio is generated.

The text-to-speech cluster carries the unit most directly. LMNT sells character allowances (200K on the $10 Indie tier, 5.7M on the $199 Premium) with per-1,000-character overage that falls from $0.05 to $0.035 as tiers climb; Hume AI ladders its Octave TTS from a 10K free allowance to 10M characters on the $500 Business tier, with overage descending from $0.15 to $0.05 per 1,000; PlayHT sold roughly 3M characters a year on its Creator plan with $4-per-10,000 overage until Meta’s July 2025 acquisition wound the product down. ElevenLabs and Hedra meter the same unit one abstraction up, through credits — Hedra’s wallet prices speech at exactly 15 credits per 1,000 characters alongside per-second video.

The API end of the cluster quotes the character in bulk. MiniMax prices its Speech 2.8 model at $60 per 1M characters (turbo) or $100 per 1M (HD); xAI’s Grok text-to-speech runs $15 per 1M characters even though its language API bills tokens; and Sarvam AI, India’s sovereign-model vendor, quotes its Bulbul TTS in rupees at ₹15 per 10,000 characters (v2) or ₹30 (v3). Speechmatics sits across both worlds, billing its TTS per character beside per-hour speech-to-text.

Beyond speech, the character is also the translation meter. DeepL bills its translation API per source character — $25.00 per 1,000,000 on the classic API Pro model — even as its consumer apps sell by the seat, and Unbabel’s Widn.ai caps free translations at 1,500 characters each. On the writing side, Rytr gates its free tier at 10,000 characters a month before flat “unlimited” plans take the meter away entirely.

How it works

There are two dominant structures. The subscription-plus-overage shape (bill = tier fee + max(0, characters − included) × rate per 1k) rules the self-serve TTS tier ladders; the flat per-million-character rate shape (bill = characters ÷ 1,000,000 × rate) rules the developer APIs. The levers:

Worked example — pricing an audiobook. A 90,000-word manuscript is roughly 500,000 characters. On LMNT Pro ($49, 1.25M included) it fits inside the allowance — effective cost $49. On Hume AI Starter ($3, 30K included) the same job runs ~470K characters of overage at $0.15/1k ≈ $73.50, while on Hume’s $70 Pro tier it fits in the 1M allowance. The unit is identical; the tier choice moves the bill by 25x — which is why the usage-metric guide treats allowance-sizing, not the headline rate, as the real comparison.

Run that same half-million characters through the flat per-million APIs above and the bill collapses to a single multiplication — the arithmetic in the diagram. The flat APIs are far cheaper per character at scale but expose no allowance to shelter behind: every character bills from the first one, and DeepL’s monthly base fee means low-volume API users still pay a premium per effective character. You can model the credit-wallet version of this math with the ElevenLabs pricing calculator.

Companies using this

12 in-corpus companies meter characters, splitting into self-serve TTS ladders, per-million-character developer APIs, translation meters, and a single character-capped free writing tier. The table below sorts by pricing model, billing units, and free-tier availability.

Patterns observed

The defining self-serve pattern is the falling-rate ladder: every multi-tier TTS vendor prices the marginal character cheaper as the subscription grows — Hume’s overage drops 3x from Starter to Business, LMNT’s 30% from Indie to Premium. The tier ladder is a volume-discount curve wearing subscription clothing — a bigger commitment buys a cheaper unit, the same shape covered in the usage-metric guide.

A second pattern separates the self-serve tiers from the flat-rate developer APIs. Where LMNT and Hume ladder the rate, MiniMax, xAI, and DeepL publish one per-million number and let volume do the work — no allowance to size, no tier to jump. The split is a segmentation move: creators and studios want a bundled subscription with a comfort-blanket allowance, while developers want a raw per-unit rate they can multiply in a spreadsheet. Several vendors run both — Speechmatics and Sarvam expose the character as a bare API rate while ElevenLabs wraps it in a subscription — because the same meter serves two buyers who want it packaged differently.

A third pattern is that voice cloning travels free with the meter. LMNT includes unlimited clones on every tier including free, and ElevenLabs ships instant cloning from its $6 Starter, because cloning drives character volume rather than competing with it. And a fourth, quieter pattern: the pure character meter is giving way to credit wallets at the platform end. ElevenLabs denominates everything in credits that map closely to characters; Hedra prices speech, video seconds, and image megapixels through one fungible balance. The character survives as the counting rule inside the wallet even where it disappears from the price card — the mechanics the prepaid-credit guide unpacks in detail.

Counterexamples & variants

Rytr is the cleanest counterexample: it uses the character meter only to define the free tier’s edge (10,000/month, hard cap), then abolishes it — $9/month buys “unlimited generations” with the meter removed entirely. The bet is that for AI writing, predictability sells better than elasticity, the opposite conclusion from the TTS cluster. PlayHT walked into the same tension from the other side: its “Unlimited” $49 plan undercut its own character meter before Meta acquired the company and withdrew self-serve pricing altogether — a reminder that an unmetered flat tier can quietly cannibalize the per-character ladder it sits above.

DeepL is the boldest variant: it runs the character meter and a per-seat meter simultaneously on one translation engine — characters for the developer API, seats for the knowledge-worker apps ($8.74 to $57.49 per user). The wrinkle is the fixed monthly base fee ($5.49 on API Pro, $26 on Growth) that sits under the per-character usage, which makes the API cheap at scale but expensive to start — at low volume that floor can dominate the bill. The character here is a genuine usage meter, but not pure pay-as-you-go — a pattern the invoicing guide flags as the “usage product with a subscription floor.”

Hume AI shows the unit’s boundary inside one product line: batch Octave TTS bills per character, but live EVI voice sessions bill per minute — when latency and turn-taking enter, the input text stops being the cost driver and the clock takes over, the same split that defines media-minute pricing. And Sarvam AI is the multi-meter extreme: it bills LLM inference per token, speech per audio hour, and TTS/translation per 10,000 characters, all in one INR-denominated stack — so the “cost driver” flips by workload. For a voice product the character and audio-hour meters dominate and the tokens are nearly free, which is exactly the blended-cost trap the character meter alone can hide.

What this means for buyers vs vendors

For buyers

Estimate in characters, not words — English runs ~5–6 characters per word, so a “50,000-word” job is a ~300,000-character job — and ask what the meter counts: whitespace, SSML markup, and regenerations typically all bill. Then match the pricing shape to your volume. On self-serve TTS tiers, shop the allowance, not the rate: the included-character ladders are where the real price differences live, and the falling overage curves mean upgrading one tier is often cheaper than paying overage on the current one (the diagram’s Hume case above). At high, steady volume the flat developer APIs win outright — but watch for a fixed API base fee, which taxes low-volume use.

If your workload spans meters, price the dominant unit first. A voice or translation pipeline on Sarvam AI is a character-and-audio-hour bill with almost-free tokens; a live conversational agent on Hume is a per-minute bill where the character meter barely registers. Estimating from the headline character rate alone will mislead you on any product that also charges by the clock. The usage-metric guide works through how to identify which unit actually drives a mixed-meter bill.

Finally, treat the wallet as a character meter in disguise. When ElevenLabs or Hedra quote credits, convert to characters before comparing — the per-1k exchange rate is what lines a credit price up against a per-1k TTS quote. Otherwise a credit price looks cheaper or dearer than it is.

For vendors

The character is the most legible meter in audio and translation AI — buyers can count it in a text editor — so protect that legibility: publish the counting rules, keep one rate per tier, and put the volume discount in the ladder rather than in negotiation, the way LMNT does. Decide deliberately whether you are selling a self-serve subscription (ladder the rate, size the allowance) or a developer API (one flat per-million rate, no allowance games); DeepL shows you can run both, but only if each buyer sees the packaging built for them rather than a compromise between the two.

If you span media types, follow Hedra and keep the character as the speech-side exchange rate inside the credit wallet rather than inventing a new opaque unit — the prepaid-credit mechanics guide covers how to keep that exchange rate visible so buyers still trust the meter. And if part of your product is latency-bound, split the meter explicitly (characters for batch, minutes for live) the way Hume does, rather than averaging the two into a number neither workload can forecast.

Watch the base-fee trade-off. A fixed API floor turns volatile per-character revenue into a predictable stream, but it dominates the bill at low volume and filters out exactly the small developers who might have grown into scale. If you add one, surface an honest “your effective per-character cost at this volume” estimate so it doesn’t arrive as bill shock on the first invoice — the same discipline the invoicing guide recommends for any usage product with a subscription component.

Explore this theme in the knowledge graph →