Cost and latency

Cost

Every session shows what it cost, broken down by component, rather than as a single number you have to trust. The session's cost carries a currency, a total, the duration_minutes and a components list of {component, cost} entries, largest first. The component names you can see:

Component	What it is
`llm`	The language model on a voice call.
`stt`	The transcriber on a voice call.
`tts`	The voice: text spoken on a call, or a reply read aloud in a text session.
`realtime`	The speech-to-speech model on a realtime call.
`backend`	On a realtime model that hands its reasoning to a separate model, that model's tokens.
`platform`	The flat platform fee on a voice call.
`chat_llm`	The model on a chat or text turn, including the per-message fee.
`analysis`	Post-call analysis, when enabled on the agent.
`summarization`	Condensing a long conversation's history to keep it within the model's context.
`judge`	The model that grades a test-suite run against its rubric.

Only components that cost something appear, so a cascaded voice call normally shows llm, stt, tts and platform, and a realtime call shows realtime (and backend where it applies) with platform.

A component you run on your own key does not appear as a charge, because Glytos is not buying it for you.

Knowledge-base embeddings - reading your documents in and embedding a search query - are charged to the organization rather than to a session, so they do not appear in a session's breakdown.

How the number is arrived at is on Pricing and credits.

Latency

A voice call feels good or bad based on the gap between the caller finishing and the agent starting. Each voice session reports where that gap went, averaged over the call and turn by turn:

Stage	What it measures
Transcriber	Time to turn speech into text.
Model	Time to the first token of the reply.
Voice	Time to the first audio of the reply.
Endpointing	Time spent deciding the caller had finished.
Turn	The whole gap, end to end.

Knowing which stage is slow is what tells you what to change - a faster model, a different voice provider, or turn-taking that is waiting too long. Guessing at this without the breakdown usually means changing the wrong one.