Two numbers sit beside every call, and blurring them is the fastest way to misread your own bill. total_charge is the two added together, and it exists so you can see a call’s full economics in one figure. Only the platform fee is ever debited from your credit balance. Debiting the provider cost would charge you a second time for spend that already went onto your own key — which is what strict bring-your-own-key means here. See provider keys.

​
The platform fee

It is prorated by the second, not rounded up to a whole minute: a 90-second voice call is charged $0.00525. Text carries no platform fee. A text session is an idle conversation window rather than a stretch of talking, so a per-minute charge there would bill how promptly your customer replies rather than anything the agent did. A text conversation costs your workspace only what its own LLM key was charged. The rate is decided by the channel the session ran on, and nothing else. Avatar time on a video call is provider cost on your own key, on top of this fee, not part of it.

​
The fee is waived on a failed call

A call that failed keeps its provider cost — the providers ran and charged you regardless — and is charged no platform fee. A close reason counts as failed if it ends in _failed, or if it is one of: error, stale, job_shutdown, job crashed, orphaned, unknown, avatar_start_failed, invalid_dispatch_metadata, missing_dispatch_metadata, unknown_tenant, missing_published_definition, rejected, insufficient_credits. Everything else is a normal completion and is charged the fee — including a call with no close reason at all, and including transferred, which is a successful escalation rather than a failure. There are no substring heuristics: a reason with “error” in the middle of it is not a failure unless it is on this list. See close reasons.
Voicemail is an answered call. There is no answering-machine detection anywhere in the product, so an outbound call that reaches a voicemail greeting connects, runs, is recorded and is billed like any other. This applies to every outbound call, not only to batches.

​
The five metered kinds

Every provider charge on a call is one of five kinds, and each is priced from the catalog entry of the exact model that ran. A call fills either llm + stt + tts or realtime — never both, because the agent ran one pipeline or the other. Avatar sits beside whichever it ran. Two flags qualify an LLM line rather than being a kind of their own:
  • purpose is conversation or analysis. Post-call analysis prices through exactly the same path as the call, so without this its cost would fold invisibly into the call’s own LLM line.
  • priority says the line ran in the provider’s priority lane, which prices at a different rate block on the same catalog entry. It is the reason one LLM line can cost double the standard rate. See priority lane.
Noise cancellation is not metered by Talqing at all.

​
Reading cost on one call

GET /v1/calls/{session_id} returns cost: The snapshot is never re-derived from today’s catalog. A rate that changes next month does not move what an old call says it cost.
cost is null until billing has computed it. A call in progress shows no cost at all — never a partial estimate. session.billing_status says which state it is in, and only computed carries a cost object.

​
Reading cost across the workspace

GET /v1/observability returns a cost section over the range, for up to seven days at a time. GET /v1/calls/stats is the smaller question — total_spend and avg_cost_per_call over priced_calls, for a window that defaults to the last 30 days. See calls.
Agent task runs are not in any of this. A task is an agent nobody talks to; it writes no session row, so neither its runs nor its spend reach the observability endpoint or the calls list. Task spend appears on the task’s own run history, and — for a task drafting an email campaign — on that batch’s page. See running tasks.

​
The credit ledger

GET /v1/billing/credits/ledger is every movement of this region’s balance, newest first, with what caused each. GET /v1/billing/credits is the balance itself, plus what it buys at today’s rates. Every entry carries amount (signed — a usage row is negative), balance_after, and created_at. Three consequences of the ledger holding only the platform fee:
  • A text session writes no ledger row at all. Its fee is zero, and a row saying “we took nothing” is noise in a statement you have to read.
  • A failed call writes none either, for the same reason.
  • A session is debited exactly once, ever. Re-pricing a call — which is what backfilling analysis does — cannot debit it twice. The fee is a function of duration and channel alone, so a re-price could not change the amount anyway. What a re-analysis does grow is your own provider bill, on your own key.
Balances are per region, because the debit commits in the same transaction as the call’s money columns. You can hold credit in India and none in the US. Never read a balance without knowing which region it belongs to.

​
Running out

The check is balance > 0 and nothing more — no reservation, no forecast of what a call might cost. It runs when you start a web call, place an outbound call, dial a batch recipient, or take an inbound SIP call. At zero, starting a new call is refused with HTTP 402:
This workspace is out of credits. Add credits in Settings → Billing to start new calls. Calls already in progress are unaffected.
An inbound call refused this way is recorded with close_reason: insufficient_credits, and is never itself charged the fee it was refused for lacking. A call already running is never interrupted. Because of that the balance can end a call slightly negative from that call’s own fee, and the next top-up pays the debt off first. An outbound batch that runs out pauses without failing and without spending an attempt, and resumes on its own when you top up. See pricing and credits for packs, checkout and the low-balance warning.

​
In the dashboard

The call detail page has a Cost section: a proportion bar of where the provider cost went, then one row per priced line — kind, provider, model, the exact quantities and the rate each was charged at — with priority and analysis tagged on the lines they apply to. Under them sit provider cost, the platform fee with its formula ($0.0035/min × 4m 12s, or waived — the call failed), and the total. Money is shown to six decimal places throughout, because a real line can cost $0.000762 and anything shorter rounds it to nothing.

Pricing and credits

Packs, checkout, the signup grant and the zero-balance rules.

Provider keys

Which providers you bring a key for, and what fails without one.