Pages
Cost Center
c105 — the qdash /cost page re-drawn in the Organization / Inventory listing vocabulary · served snapshot 2026-10-01
Period
Every $ is a list-price estimate — this seed reports no billed cost. Click a row to name the page it opens on the Fable app.
Cost Center
What AI costs, who and what it is attributed to, and where it is heading — billed where the provider reports it, list-price estimate otherwise.
Spend
Usage at list price · Fleet · 30 days
$
1926
$909Payments
$404Data
$324Platform
289Other
$66 a day · $186 metered (billed), $1,740 on plan seats (at list)
Pace
Q3 AI spend ceiling · trailing 30 days
96
%
$1,921Spent
$79Left
$1,921 of $2,000 · the $25/day guardrail tripped on 22 of 29 days
Rate
Blended $ per 1M tokens, cache priced in
$
16
/M
claude-sonnet-4-647 %
$20/M
claude-opus-4-838 %
$94/M
gpt-5.5-codex8 %
$4.92/M
claude-haiku-4-55 %
$6.56/M
claude-opus-4-8: 6 % of tokens, 38 % of spend
Spend by
Team | Spend | Share | Tokens | $ / 1M | Responses | Spend trend |
|---|---|---|---|---|---|---|
Payments 4 people | $909 | 47 % | 53M | $17 | 1,085 | +19% |
Data 3 people | $404 | 21 % | 27M | $15 | 730 | -30% |
Platform 4 people | $324 | 17 % | 25M | $13 | 931 | -51% |
Frontend 3 people | $289 | 15 % | 17M | $17 | 386 | +24% |
Which department does this charge back to? 4 teams · assigned per person on Users. Spend trend: each team's daily cost, later half vs earlier half — the only lens with a served day series.
Costliest sessions
User | Harness | Model | Repo | Started | Duration | Tokens | Cost |
|---|---|---|---|---|---|---|---|
| tomas Payments | Claude Code | claude-sonnet-4-6 | payments-core | 06-17 16:31 | 22m | 1.4M | $37.33 |
| elena Payments | Claude Code | claude-haiku-4-5 | payments-core | 06-26 13:02 | 30m | 1.1M | $35.69 |
| tomas Payments | Claude Code | claude-sonnet-4-6 | payments-core | 06-24 12:30 | 38m | 947k | $32.63 |
| omar Frontend | Claude Code | claude-sonnet-4-6 | web-checkout | 06-08 14:59 | 39m | 651k | $27.67 |
| tomas Payments | Claude Code | claude-sonnet-4-6 | payments-core | 06-23 13:33 | 7m | 772k | $25.93 |
| priya Payments | Claude Code | claude-sonnet-4-6 | payments-migration | 06-18 10:57 | 33m | 409k | $25.65 |
| priya Payments | Claude Code | claude-opus-4-8 | payments-migration | 06-22 11:37 | 35m | 378k | $25.53 |
| marcus Payments | Claude Code | claude-sonnet-4-6 | payments-core | 06-08 13:48 | 34m | 688k | $25.46 |
| elena Payments | Claude Code | claude-sonnet-4-6 | payments-core | 07-01 13:48 | 33m | 1.2M | $23.63 |
| tomas Payments | Claude Code | claude-sonnet-4-6 | payments-core | 06-18 12:07 | 12m | 760k | $22.34 |
A receipt, not a target: the session behind a spike, opened in Activity. Ranked from the 200 sessions /api/sessions serves for the period (capped at 200).
Ceilings
Ceiling | Limit | Current | State | Reading |
|---|---|---|---|---|
Daily spend over budget Control · fleet · per day | $25 / day | $66 last 24 h | over | over on 22 of 29 days — set below a typical day ($66) |
Q3 AI spend ceiling Target · fleet · trailing 30 d | $2,000 | $1,921 · 96% | near | served as $0 — the target's window misses the data |
Q3 token budget Target · fleet · trailing 30 d | 150M | 122M · 81% | under | excluding cache — 1.4B cache reads are not counted |
The guardrail is a control (Policies), the targets are campaigns (Campaigns) — set and changed there; read here against the same trailing-30-day numbers the page shows.
What changed from today's /cost
- Four StatTiles (total / provider-reported / estimated / cache %) → three DataMachine cards that each answer a question: how much and who (Spend), where it is heading (Pace), what we pay per unit and why (Rate).
- Two BarList panels + four titled tables → one ledger pivoted by lens with a column picker, the same columns under every lens.
- The Spend guardrail panel → the Pace card + a Ceilings table: the control and both campaign targets read against the same trailing-30-day numbers.
- The spend-over-time chart → the Spend card's spark plus per-team trend sparks in the ledger (the c103 Teams pattern).
- New: Repo, Account and Token-kind lenses, costliest-session receipts, Plan vs Metered and Company vs Personal on accounts.
- Round 2: plan-seat dollars read as usage at list price (Spend card, Account lens) until seat prices exist; Ceilings is one self-contained block (CostCeilings) that can move to its own page; every row names the Fable page it opens.
What the snapshot says
- Cache is 62 % of spend and also the biggest saving: cache writes (38 %) out-cost output (26 %), while cached reads saved ≈ 2× the whole bill against the fresh-input rate.
- claude-opus-4-8 is 6 % of tokens and 38 % of spend at ≈ $94 / 1M, against ≈ $20 for sonnet — model choice is the lever; Payments' migration work is opus-heavy.
- The trailing 30 days sit at 96 % of the $2,000 ceiling, yet the campaign serves current = $0 and "on track": its window is wall-clock and the seed's data ends 2026-07-01.
- The $25/day guardrail tripped on 22 of 29 days in the 30-day view: set below a typical day, it is noise rather than a signal.
- 90 % of spend is on sign-in (plan) accounts — that $ is a list-price equivalent, not an invoice. Only the API-key share is what you are billed per token.
Knowable, not yet served
- Per-day cost by user, model, provider, harness or repo — only fleet and team series are served (a group_by on /api/costs).
- Token kinds per team — /api/models and /api/providers ignore ?team=.
- Provider-billed cost — provider_cost is 0 across this seed; every row is estimated.
- All sessions in a long window — /api/sessions caps at 200, so receipts past ~2 weeks rank a sample.
Not knowable from the wire
- What a plan seat actually costs (Claude Max, Cursor, ChatGPT Team) — needs a price per seat entered by the customer, then the page can show seat cost vs list-price usage.
- Output or value per dollar — the sensor sees sessions, tool calls and repos, not merged work; Repo is the honest proxy.
- Department beyond team — the CFO's cost centre needs a team → department map (an Organization setting).