Pricing and credits
Glytos bills what a conversation actually consumed. There are two meters, one per channel, and both charge each component in the unit its provider sells.
Voice calls
A call's charge is the sum of four components:
| Component | Metered on |
|---|---|
| Language model | The tokens the model actually used, at that model's list rate. |
| Voice | The characters actually spoken, at that voice's per-character rate. |
| Transcriber | The audio actually sent to the transcriber, measured in seconds. For a streaming transcriber that is the whole call; for one that transcribes each spoken segment separately, it is only those segments. |
| Platform fee | Call minutes, at a flat rate. A call holds a worker slot and a media session for its whole duration whether or not anyone is speaking. |
The margin applies to provider cost only, never to the platform fee.
This is why two one-minute calls can cost different amounts. A minute of rapid back-and-forth genuinely burns several times the tokens of a minute spent waiting on one slow answer, and the bill reflects that rather than averaging the two.
If a provider reports no usage for a component - a speech service reached over its own HTTP endpoint, for instance - that component is estimated from the call's duration instead.
Realtime agents
An agent on the realtime engine has one model doing all three jobs, so it is charged as one component plus the platform fee, not four. It is metered on the audio tokens the model reported, at that model's own published rate.
A full-duplex realtime model that hands its reasoning to a separate model is the exception: it is two products on one call, so it bills two components plus the platform fee - the live model in its own unit, and the reasoning model on the tokens it used.
Audio tokens are an order of magnitude dearer than text ones, and providers discount a repeated prompt prefix far more steeply than text models do - so a realtime minute is priced on a different scale from a cascaded minute and the two are not comparable. Where a provider sells realtime by the minute instead of by the token, that is what you are charged. Run one real call and read its cost breakdown before committing to realtime for volume.
Chat and text
Text turns are metered per token, from the counts the provider itself reports, priced at the model's list rate plus a margin and a small flat fee per message.
Two things are never billed: a step that speaks a fixed line without calling a model, and a reply that came back empty.
Post-call analysis
Analysis is metered per token on the model that ran it, plus a margin and a small flat fee, and appears on the session as its own component. An analysis that produced no result is not charged.
Knowledge base
Embedding your documents when they are added, and embedding each search query, is charged to the organization on the tokens used. It is not tied to a session, so it does not appear in a session's breakdown.
Bring your own keys
When a provider runs on your own key, Glytos is not buying that provider for you, so it is not charged. The flat platform fee still applies.
Seeing what a call cost
Every session shows its cost broken down by component alongside the transcript, so a charge is auditable rather than a single number. See Cost and latency for the component names.
Credit
An organization runs on prepaid credit. When you sign up, your first organization receives a one-time welcome credit so you can build and test before paying anything. Where email verification is on, the credit is granted once you verify your address. Organizations you create later start with no credit.
Running low
Credit cannot go meaningfully negative, and calls are bounded rather than cut off mid-sentence without warning:
- A call is refused if the balance is below the minimum needed to start one.
- A call ends at the earliest of: the agent's own maximum duration, the platform's hard cap, and the point the remaining balance can pay for.
- Chat checks the balance before every turn.
This means a nearly-empty account gets a short call, not an all-night one. It is a guard against a runaway agent, not a way to bill you past your balance.
Topping up
Top up from Settings - Billing. Two things are required first:
- a verified email address, and
- a billing profile - your name, email and country are required; a phone number is optional. The country decides the currency you are charged in (below).
Invoice details are only asked for if you tick that you want an invoice, and are only stored in that case. An invoice needs an address and a city. A company also needs its company name, and a company in Turkey its tax office and a valid tax number; an individual in Turkey needs a valid national identity number. An identity number is encrypted at rest and is never returned by the API after it is saved - reads show it masked.
Invoices are not generated automatically yet. Ticking the box records the details an invoice needs; ask support if you need the document itself.
Your credit balance is held in one currency. What you are charged follows the country on your billing profile: customers in Turkey are charged in lira at the central-bank rate plus a small conversion margin, everyone else in the balance currency with no conversion. The checkout shows the exact amount before you commit.
This is not a preference setting. A Turkish-issued card cannot pay a foreign currency at all, and the currency is fixed before any card is seen.
If you pay and close the tab before returning, the payment is still credited: the result is reconciled from the provider afterwards, and crediting the same payment twice is impossible.