Picture a small restaurant, invented for this example. On Thursday the kitchen decides the weekend special will be braised short ribs. On Friday morning the owner updates the fact in the agent: "Weekend special: braised short ribs with polenta, 24 dollars." On Friday at noon the first caller dials in and asks, "Do you have a special this weekend?"
Everything below is an illustration of how the Speakvora pay-once agent is built to behave, with invented numbers. The arithmetic is mine; the mechanics are from the agent's code.
The first time
The agent has recorded lines for the things a restaurant says every day: the greeting, hours, how to book. None of those is the special. So the answer to this caller is a new sentence, written from the owner's facts by the connected AI, say: "This weekend we have braised short ribs with polenta for 24 dollars." I counted that: 66 characters.
The agent checks the store for those exact words and finds nothing. If voicing is allowed, it sends the sentence to the voice engine, stores the audio under a name derived from the exact text, and plays it. In some paths, such as the greeting, a line that is not ready yet is covered by the phone carrier's built-in voice while it is voiced in the background, so the next caller gets the recording. That is the honest first-time cost: a caller may hear a different voice for that sentence, and the agent logs when it falls back.
Cost, in the default cloud mode: 66 characters. At Speakvora's $3 per 1M characters that is 66 / 1,000,000 x $3, about $0.0002. The agent's guard counts those 66 against its monthly budget of 100,000 characters, which is 0.066 percent. (Per Speakvora's changelog of 21 November 2026, speech made by a pay-once agent is not billed per character; the budget is a safety limit, not a meter.)
The second through fiftieth time
By Saturday evening, say fifty callers have asked. Caller two asks the same thing, and the AI produces the same sentence. Same words, same key, found in the store: the stored clip plays, with no voicing and no characters spent. The agent refreshes that line's last-used time each time it plays. Callers three to fifty are the same. Total cost for fifty callers: still 66 characters. If you had voiced each reply live, it would be 50 x 66 = 3,300 characters.
Honesty about "the same words": this works only when the AI's sentence is identical. If it produces "This weekend's special is braised short ribs with polenta, 24 dollars" on call seven, that is a different string, so it is a new line, voiced once, costing about 70 characters. The agent has a learning feature that counts repeated customer-facing lines and, at the repeat count the owner sets (default 3), turns a repeated line into a library entry, but it only learns generic lines. Anything with digits is skipped, and this sentence has a 24 in it. So this special would not be learned into the library; it relies on the cache of recently voiced lines, which is exactly what the cache is for. In practice a few wording variants will exist by Sunday, each voiced once. Say five variants: 5 x 70 = 350 characters for the whole weekend.
When the special changes
The next Friday the owner types a new special. Last week's sentences are still in the store, and nobody asks for them. Ordinary voiced lines are kept for 30 days since their last play, so they expire on their own around a month later. Nothing for the owner to clean up, and no stale audio plays, because the key is the words and the new special has new words.
Over a year, with a new special every week and, generously, five wording variants a week at 70 characters, the cost is 52 x 5 x 70 = 18,200 characters. That is 18.2 percent of one month's free allowance, spread across twelve months. Even if the owner announces a second offer every week, you are around 36,000 characters a year.
What would change this
A busier restaurant changes nothing about the arithmetic, since callers add plays, not characters. What does change it: a long, wordy special description (a 300-character sentence costs four and a half times as much to voice once), a language model that rewords freely (more variants), or a budget that is already used up when the new special arrives, in which case the carrier's voice covers it until the budget renews, and the log says so. If the restaurant would rather never hear the carrier voice on the special, the answer is to keep a spare budget, or to pre-voice the special on Friday morning before the first caller. In the agent's dashboard, lines that are added or edited are voiced right away, with the code noting that a call then plays them instantly.
The practical advice for the owner is short. Put the special in the facts when you decide it, test-call once on Friday so the first real caller does not become the person who triggers the voicing, and let the agent forget last week's.
For more on how the pay-once agent works for restaurants, see speakvora.com; it covers what is included and what the optional support plan costs.
Method and sources
All numbers are for an invented restaurant and are labelled illustrative. Character counts (66, and about 70 for variants) were counted by me. The $3 per 1M characters price and the 100,000-character free allowance are from the Speakvora pricing page, 22 November 2026. The not-billed-per-character statement is from the Speakvora changelog of 21 November 2026. Behavior (cache keyed by exact words, 30-day retention with refresh on use, background voicing with carrier-voice fallback and logging, default 100,000-character budget, learning at a default repeat count of 3 that skips lines with digits, voicing of added or edited lines from the dashboard) is from the agent source files audio.py, engine.py, learned.py, guardrails.py and dashboard.py, read on 22 November 2026.
Top comments (0)