The agent calls Anthropic's Claude API for every reply. Each call has a token-based cost. BandBrain logs every call and attributes it to your tenant + the person who triggered it.
Where to see it
/dashboard/usage — shows per-tenant Claude spend over the last 30 days:
- Total — micro-cents summed
- By surface — sms / email / web / discover / extract_day_facts / etc.
- By model — typically all sonnet now; opus and haiku are deprecated for BandBrain
- By person — which roster member's queries cost the most
- By day — daily breakdown
What costs what
| Surface | Typical cost per turn | Notes |
|---|---|---|
| SMS reply | $0.005 – $0.02 | Single turn, prompt cached |
| Web chat | $0.005 – $0.02 | Same |
| Day-facts extraction | $0.01 – $0.05 | Runs after each file edit; debounced + incremental |
| Discovery (new file) | $0.05 – $0.20 | One-time per file |
| Indexing (embedding) | $0.001 – $0.005 | Gemini, not Claude — not in this dashboard |
| Memory extraction | $0.001 – $0.005 | Best-effort, post-turn |
| Comms automation run | $0.01 – $0.05 | Body composition uses one agent turn per fire |
Why it's cheap
- Prompt caching keeps the static prefix (tools, system prompt, recent context) at 0.1x cost across consecutive turns within ~5 minutes
- Sonnet, not Opus — defaults are tuned to Sonnet ($3/$15 per MTok) which handles tour-data tasks fine
- Per-source-path incremental extraction — editing one tab doesn't re-extract the whole sheet
- Debounced jobs — burst saves get collapsed into one extraction
Spend depends on how much your band uses BandBrain: more messages, forwarded mail, and automations mean more AI calls.
When to worry
- Day-facts extraction running on every file change for a folder with hundreds of files
- A runaway job retry loop reporting errors to Sentry (separate cost)
- Opus accidentally getting selected (we've removed Opus from the BandBrain code paths; if you see it on the usage page, something's wrong)
The 30-day chart on the usage page makes spikes visible quickly. Day-over-day usually stays flat unless you're actively backfilling or onboarding lots of new files.