How we calculate
Every number on this site is computed from a vendor's official pricing page. This page explains the arithmetic, its limits, and how the site is paid for.
Where the numbers come from
We read each vendor's own pricing page and documentation, never third-party summaries. Each pricing page on this site lists its sources, the date we read them, and the exact quoted text the numbers come from. If a vendor's page is unclear, we say so instead of guessing.
Every price on the site, with the page it came from and the date we read it, is also a spreadsheet you can download from the comparisons page.
Turning credits into minutes
Most AI tools sell credits or characters, which tell you little about what you get. We convert them into units you can compare: minutes of speech.
- Where a vendor states its own conversion, we use it. ElevenLabs, for example, equates about 1,000 characters with a minute of speech; every vendor's figure is on characters per minute of speech.
- Cost per unit assumes you use the plan's whole monthly allowance. If you use half, you pay twice as much per minute.
- Voices billed in tokens, like Google's Gemini TTS, are converted with the vendor's own audio tokens per second, plus the text at the vendor's own characters per token; see Gemini TTS pricing. Where a vendor states no such rate, as for OpenAI's gpt-4o-mini-tts, we show its token prices and don't convert them.
- Fish Audio's API bills UTF-8 bytes of text, which is one byte a character in plain English; we count it per character, so text with accented or non-Latin letters costs more than shown.
- Prices are in US dollars and exclude tax, as the vendors list them.
- Monthly costs use monthly billing where a plan offers it, otherwise its yearly price per month, with free allowances included. "Bigger plan" means a plan's minutes run out at that volume and no extra-minute price is published.
- Head-to-heads take each side's cheapest option for the volume, and bold marks the cheaper. Comparisons are on price only: voices, languages, latency and licences differ, so try both.
- "What the numbers say" on a vendor's page sets its cheapest option at each volume against every vendor selling the same kind of product (an API, or an editor app), and its lowest rate a minute against the median of every option of that kind. Ranking tables there show the rows either side of the vendor; the calculators and hub pages show the whole field.
- Promotions are shown separately, with their end dates, and never used as the headline price. A lower price that the vendor ends on a set date, like Gemini 3.8 TTS's introductory price, is shown beside the regular price, which is the one we rank.
Comparing speech-to-text prices
Transcription is sold per second, minute or hour of audio, or in tokens, so we convert every price to dollars per hour of audio and per minute. The rules:
- We compare regular pay-as-you-go list prices. Volume tiers that apply automatically, like Google's, are counted; commitments and prepaid plans are listed on each vendor's page but not ranked.
- Free minutes count only when they recur every month without expiring. One-off sign-up credits, first-year allowances and separate free tiers are listed, not subtracted.
- Promotions are shown with their end dates, or with a note that none is given, and never used to rank. A model with only a promotional price is listed but not ranked.
- Token-billed models get the vendor's own per-minute estimate, or one computed from the vendor's stated tokens per second, and are marked as estimates.
- Medical, custom-trained, human and dictation products are priced on each vendor's page but kept out of the rankings, because they do a different job.
- Recorded audio (batch) and live audio (streaming) are ranked separately. Where a vendor's price differs by region, we use the region named on the page.
- We don't rate accuracy, speed or language support. Every number comes with the quoted text and address of the page it came from.
Comparing AI voice agent prices
Voice agent platforms charge per minute of conversation, but they split the bill differently. We keep each vendor's own split rather than forcing one shape on all of them:
- The platform's own fee is shown apart from what it passes through: speech to text, the language model, text to speech and the phone line, each at the price the vendor's page gives.
- Where a vendor gives a range for a provider, as Vapi's calculator does, we show the range and carry it into every total.
- Where a vendor bills a part by the token and gives no per-minute price, such as the language model on ElevenLabs Agents, we don't convert it. The calculator asks for your own per-minute figure and labels it as yours.
- OpenAI's Realtime API is priced per minute only with OpenAI's own statement of audio tokens per second, and only as a floor: each second of audio counted once. You set how much of the call the agent speaks.
- Cheapest and dearest picks take the lowest and highest priced option in each choice the vendor lists, without add-ons. Phone lines are compared separately; where a vendor leaves the line to your own carrier, we show Twilio's US list prices.
- Every price in the comparison carries the exact text it was read from, and the build stops if a number no longer matches its quote. Each platform's page shows the quotes; head-to-heads link to them.
- Monthly tables take each side's cheapest picks, then its dearest, on whichever plan costs least for the volume; plans with included minutes or a lower rate count where they lower the bill. Phone number rental, extra concurrency and taxes are left out. The calculators open on an inbound call to a Twilio US local number at Twilio's price; set the line to none for web or app calls.
- Latency figures are the vendors' own claims. None says what is timed or under what conditions, and they don't use the same measure, so they can't be ranked against each other; test on your own calls.
How often we re-check
Tracked pricing pages are re-checked on a schedule. When something changes we update the numbers, the "checked" date on the page, and the price change log.
What we don't do
- We don't claim hands-on testing we haven't done. When we run our own measurements, the page says how.
- We don't sell placements. Rankings come only from the published numbers, and tools that pay us nothing are included.
- We don't invent reviewers or personas.
How we make money
Links to vendors can be affiliate links: once a vendor's programme accepts the site, it may pay us a commission if you sign up through one, at no extra cost to you. Those links are labelled "Affiliate link". No programme has accepted the site yet, so today no link here earns anything. Commissions never change our numbers or the order of results; the calculator sorts by cost alone.
How the site is made
Data collection and checks are automated. Every figure is computed by code from the cited source data rather than typed by hand, which is how we keep hundreds of numbers consistent.
Corrections
If you spot a wrong number, see the about page for how to reach us. We fix errors and log the correction.