Cost estimates and observed cost

Status: Current | Last reviewed: 2026-09-02

TrueServe can forecast provider cost while you design an agent or workflow and show metered cost after it runs. These are operating estimates, not invoices or prices charged to your customers.

Estimate an agent

Open an agent, choose Settings, and find Cost estimate.

Choose Edit assumptions to change the turn counts, input/output tokens, speech ratios, provider/model audio tokens per second, caller and agent words per spoken minute, text tokens per spoken word, tool schema/call/result sizes, transcript copies, session length, history replay interval, or characters per word for this agent. Save, then refresh the estimate. These overrides do not change another agent.

For token-priced realtime audio, TrueServe calculates projected tokens as active audio seconds × the selected provider/model token cadence. Caller and agent audio are calculated separately. OpenAI Realtime defaults to 10 input and 20 output audio tokens per second, or 600 input and 1,200 output tokens per active audio minute. Nova Sonic has its own 25 input and 50 output tokens per second planning profile. Words per minute affect text transcripts and character-priced services only; changing WPM does not change audio tokens. Nova text estimates also include the configured prompt, enabled tool definitions, expected tool calls and results, speech transcripts, and growing conversation history replayed during session continuation. These are planning assumptions; provider-reported speech and text token usage takes over after the conversation runs.

Some capability toggles register several callable tools. The estimate expands those capabilities to their runtime tool-schema count rather than counting one toggle as one tool. Expand Typical component breakdown to see the resulting session-context, history, tool-result, transcript, and tool-argument quantities. Costs below $0.10 display up to six decimal places so a per-minute value remains consistent with its hourly projection.

ElevenLabs Text to Dialogue estimates use the selected model's character rate. v3 Conversational is $0.05 per 1,000 characters; the standard Eleven v3 model remains $0.10 per 1,000 characters. Observed cost uses the actual character count recorded by the TTS stage.

Estimate a workflow

Open the workflow settings sheet. Cost estimate shows low, typical, and high cost per execution. A batch workflow also shows per-batch and per-item values.

Workflow assumptions include input items, loop count, retry probability, tool iterations, LLM input/output tokens, and batch size. Branches and loops are projections; the notes call out equal-probability branches or bounded cycles.

Publishing an agent or workflow stores the estimate with that version. Later price changes do not rewrite the saved publish record.

See actual metered cost

The value appears as provider usage arrives. Choose refresh if the execution has just finished. “Not available yet” means no usage rows were recorded; “partial pricing coverage” means usage exists but at least one meter has no matching rate. When the execution matches a published agent/workflow version, the card also shows whether observed cost is inside the saved low/high range and its variance from the saved typical estimate. Expand the card to inspect individual meters.

Maintain prices (administrators)

Go to Settings → Tenant Settings → Costing.

  1. Choose a tenant contract or, for platform administrators, the public-retail fallback.
  2. Load the reviewed public catalog, import normalized contract JSON, or use the AWS Price List import for an AWS-only draft.
  3. Review the draft, its effective date, rate count, and coverage.
  4. Publish it. Drafts never affect estimates.

An unchanged scheduled import re-verifies the current version. A changed import shows how many selectors were added, changed, or removed before publication. Expand Unpriced catalog selectors to see the exact service, provider, model, and meter behind each coverage gap.

Tenant contract rates override public rates. The coverage cards show how many enabled provider/model meters can be priced. Fix gaps before treating forecasts as complete. A “Review due” state means a source needs verification even if its rates still calculate.

The reviewed public catalog includes enabled OpenAI chat, Responses, realtime, transcription, translation, and TTS models; direct Claude models; supported AWS Bedrock and Transcribe models; and ElevenLabs TTS, Scribe, Speech Engine, and Voice Changer. Token prices and workload assumptions remain separate: for example, Nova 2 Sonic keeps AWS's exact per-speech-token rates while the editable speech assumptions project how many tokens a minute contains. gpt-4o-mini-tts uses a documented planning conversion from its text/audio token rates. These approximations appear in the estimate notes and observed provider usage replaces them after execution.

Use Reconcile provider invoice to enter a provider, matching date window, and the provider’s billed total. TrueServe records the difference between that amount and metered cost. Large differences usually mean a missing model/region rate, cache tier, billing minimum, or shared provider endpoint that lacks tenant allocation.

Maintain workload assumptions

The same Costing screen manages effective-dated workload defaults separately from prices. Choose tenant scope, or platform scope if you are a platform administrator, review the low/typical/high JSON, create a draft, then publish it. Existing published estimates and explicit agent/workflow overrides remain pinned to their original versions.

Usage calibration analyzes the last 30 days of measured quantities. Once a metric has enough samples, it proposes p10, median, and p90 values for the low, typical, and high scenarios and shows the change from current assumptions. Choose Review proposal in editor, inspect the full draft, and publish only after the workload owner approves it. TrueServe never auto-publishes a calibration proposal.

Reading the scenarios

Low, typical, and high are independent workload assumptions—not confidence intervals. A total of zero is valid only with complete coverage and genuinely zero projected usage. If coverage is partial, unknown components are excluded from the displayed total and listed in the notes.