Why our free tier has an unlimited engine

Two kinds of voice engine

Every voice engine on our roster does the same job, but they live in two different economies. The proprietary engines are served by their vendors, who charge for every character of inference because every character runs on hardware they pay for. Those rates are real costs, and pricing that passes them through is fair.

Open-weight engines broke that pattern. When a lab publishes a model's weights under a license like Apache-2.0, anyone can host it, and hosts start competing on price for identical output. Kokoro is the open-weight engine on our roster: a lightweight model, published openly, served at a rate that would have sounded like a typo two years ago.

Most pricing pages were designed for the first economy. Ours had to answer a more interesting question: what should a plan look like when one engine in the roster is, for practical purposes, too cheap to meter?

What generation actually costs

Here are the raw engine rates, straight from the same registry our studio reads. We publish them in the open, on the engines page and here, so you can check that our routing and pricing are fair rather than taking our word for it.

The arithmetic of unlimited

Look at the Kokoro row: $0.62/M in · $0 out. A million characters of input is, by the house math of about 1,000 characters per minute, roughly 17 hours of finished audio. At that rate, even a ferocious drafting month, the kind where you regenerate every paragraph five times, costs less to serve than the coffee you drank while doing it.

When the marginal cost of a retake rounds toward zero, charging per retake stops being cost recovery and becomes something else: a meter kept running because meters are profitable, not because the electricity bill demands it.

So we did the obvious thing that pricing pages rarely do: Kokoro is unlimited on every plan, including Free. It never touches your monthly allowance. Draft as much as you want; the arithmetic genuinely does not care.

Per-character anxiety is a choice, not physics

Once one engine is effectively free to serve, an all-engines character meter is revealed as a design decision. A platform that meters every character earns a little more from every retake, so it has no reason to tell you that drafting could cost nothing. We would rather earn renewals than retakes.

Honesty cuts both ways, though, and here is the other half. The premium engines on the table above carry real per-character costs that someone has to pay, and on paid plans that someone is us, inside a flat price. So premium generation draws from a monthly allowance, sized in plain time on the pricing page, and the studio shows you the exact character count before a big job runs. Flat does not mean infinite; it means the meter is a budget you can see, not a tax on iteration.

It is also fair to say what unlimited does not buy you: Kokoro sits at 1060.25 on the third-party listener arena, below the premium engines, and we print that number rather than hiding it. It is the fastest, cleanest draft voice we have, not the most expressive finisher, and the plan treats it as exactly that.

How the two lanes work together

The free plan is the whole philosophy in miniature: unlimited Kokoro drafting, plus 10,000 premium characters a month, about 10 minutes, to hear what the expressive engines do with your script. Draft on the unlimited lane until the words are right, then spend allowance on the keeper take.

That loop, draft free and finish premium, is the workflow the studio itself suggests, and it exists because the economics above make it possible. The engine-by-engine reasoning lives in the engines guide, the metering details in usage and limits, and the draft engine has a page of its own.

None of this required generosity. It required noticing that the cost structure changed and refusing to keep charging as if it had not.