Skip to content

Usage & cost accounting

The dashboard's Usage tab answers three questions: how many tokens flow through each surface, who they're attributed to, and what it costs. This page explains the two dollar figures, how attribution works, and how to adjust the price table.

Two dollar figures: Billed vs Equivalent value

Every recorded turn (and every metered side-call) gets priced. There are two numbers, and they mean different things:

  • Billed — what actually hits a card. Only surfaces backed by a metered API key contribute: text-to-speech, transcription, and the moderation/labeler model calls, plus any reasoning engine you run in API-key mode. Claude on a Max/Pro subscription contributes $0 here.
  • Equivalent value — what the same usage would cost at public API list prices, computed for everything — including subscription engines. This lets you see the value you're getting from a flat subscription even though you aren't charged per token for it.

So a box that runs Claude on a subscription and uses OpenAI only for voice will show a large Equivalent value and a small Billed total (just the voice + moderation spend).

The split is driven by a billed flag decided at record time from the engine/provider's auth mode (subscription → not billed; api_key → billed). Both numbers ride the shared event contract (cost_usd = equivalent, billed_cost_usd = billed) into the collector and the Usage view.

Attribution: by user, not by raw handle

The per-user table groups by the trust principal, not the raw channel identity. A person with several linked handles in the Access tab — a Discord id, an email, a GitHub login, an MCP client id — folds into a single row. The resolution happens at display time against the trust store's identifier map, so it respects Access-tab edits live (link two handles and their history merges on the next refresh, with no re-processing). Unlinked identities fall back to a case-normalized key, so casing variants (e.g. moneo/Moneo) still merge before you formally link them.

Under the hood the token-usage event is attributed to the raw sender (matching the inbound message), and also carries the resolved principal in its payload — so the raw fidelity is preserved while the view folds up to the person.

Token estimation fallback

Some engines don't report token counts (e.g. a Gemini turn whose stream omits the usage line). Rather than log zero tokens for a real turn, asmltr estimates from text length (~4 chars/token) and flags the event estimated. The input estimate is a floor (it counts the user's text, not the full system prompt/history), so treat estimated rows as a lower bound.

Metered spend breakdown

The Metered spend · by feature & provider panel breaks the Billed total down into the side-surfaces that generate it:

Feature Priced by Typical provider
Text-to-speech characters synthesized OpenAI, ElevenLabs
Transcription (STT) seconds of audio OpenAI
Moderation model tokens OpenAI (or Anthropic)

These are emitted as token-usage events tagged with a feature in their payload (via shared/usage.js auxUsage()), then rolled up per (feature, provider, model, units) in the collector's usage_aux table and returned on GET /api/usage as aux[].

STT duration is estimated

Most transcription models don't return an audio duration, so seconds are estimated from the encoded clip size using a nominal bitrate per container (see estimateAudioSeconds). STT is cheap ($/min), so a ballpark keeps the Billed total honest without decoding the media.

The price table & overrides

Prices live in shared/pricing.js as a best-effort snapshot of public list prices:

  • models — USD per 1,000,000 tokens, { in, out }. Matched by longest prefix on the model id, so gpt-4o-mini-2026-… resolves to gpt-4o-mini (not gpt-4o), and a trailing date suffix is ignored.
  • tts — USD per 1,000 characters. Any unknown eleven_* model falls back to a generic ElevenLabs rate.
  • stt — USD per minute of audio.

List prices drift, so the whole table is overridable — drop a JSON file at ~/.asmltr/pricing.json (or point ASMLTR_PRICING_FILE at one) and it's deep-merged over the defaults:

{
  "models": { "opus": { "in": 15, "out": 75 } },
  "tts":    { "eleven_turbo_v2_5": 0.15 },
  "stt":    { "gpt-4o-transcribe": 0.006 }
}

Changes apply on the next process start (or call pricing.reload()).

Data model (reference)

  • events.cost_usd / events.billed_cost_usd — per-event equivalent + billed value (append-only spine).
  • usage_rollup — per (bucket_hour, surface, identity) token + cost rollup (the per-user table).
  • usage_aux — per (bucket_hour, surface, feature, provider, model, units) metered-spend rollup (the breakdown panel).
  • GET /api/usage?since=<ms> → { usage: [...], aux: [...] }.