Source: qdash /cost — replica + c66 currency ⇄ tokens unit switch (data snapshot 2026-08-14)
AI spend by model, agent, and user. Unbounded consumption is OWASP LLM10.
Billed plus estimated, across all providers.
Billed cost straight from the provider's usage API.
List-price math where the provider doesn't report cost (rows marked est).
Share of context served from cache instead of sent fresh.
Spend guardrail
live| Target | Current | Goal | State |
|---|---|---|---|
| Fleet spend · August | $1180.35 | $2500.00 | under |
| Batch pipeline tokens · weekly | 29.86M | 25.00M | over |
Spend by agent
Spend by user
Spend over time
Team · spend
Team | Tokens | Responses | Cost |
|---|---|---|---|
| unassigned est | 34.84M | 4721 | $1180.35 |
Model · spend
Model | Tokens | Responses | Cost |
|---|---|---|---|
| claude-fable-5 est | 3.23M | 3356 | $1077.70 |
| claude-opus-5 | 608.8k | 720 | $57.12 |
| gpt-5.6-sol est | 23.33M | 188 | $33.16 |
| gpt-5.5 est | 4.80M | 39 | $6.94 |
| claude-haiku-4-5 est | 2.17M | 385 | $4.47 |
| gpt-5.6-luna est | 481.2k | 19 | $0.63 |
| gpt-5.6-terra est | 220.6k | 10 | $0.32 |
| — est | 0 | 4 | — |
Agent · spend
Agent | Tokens | Responses | Cost |
|---|---|---|---|
| discovered est | 33.85M | 3426 | $881.03 |
| Claude Code CLI est | 668.8k | 924 | $185.16 |
| claude-cli est | 326.1k | 371 | $114.15 |
User · spend
User | Tokens | Responses | Cost |
|---|---|---|---|
| unknown est | 33.85M | 3427 | $881.03 |
| mark@qpoint.io est | 994.8k | 1294 | $299.31 |
c66 design notes — decisions & open questions
- Control: right-aligned large UxToggle flanked by "$ - Dollars" and "Tokens" labels (active side emphasized); the estimated-$ caveat tucks in beside it in $ mode. In qdash it rides the URL as ?unit= (house useUrlState pattern, no refetch). Earlier rounds: second UxTabGroup, then a small left-aligned toggle.
- Tile remap: total spend → total tokens; provider-reported / estimated (currency-only concepts) → input / output tokens — needs total/input/output_tokens added to /api/costs totals. Cache % is unit-agnostic and never moves.
- Token totals exclude cache reads/writes (tokens = input + output + reasoning) — hence the "excl. cache" sub on the tile, so 34.8M vs the ~96% cache rate reads honestly.
- Estimated-$ caveat hides in token mode — token counts are provider-reported, never estimated.
- Guardrail stays $: the budget rule is a currency control; only the Last-24h card flips which unit leads. Targets keep per-target units (metric: cost | tokens) — exempt from the toggle.
- Tables: both unit columns always visible; the toggle just moves the initial sort and the title suffix.
- Per-card captions: one-line educational text directly beneath each KPI card, swapping with the unit; token mode teaches input vs output (and the 5× price asymmetry), $ mode teaches provider-reported vs estimated. Earlier round tried a single strip under the grid.
- The batch-day spike: 08-13 is 29.9M tokens but only $143 — token mode flattens every other day. Log scale? Dual-axis overlay instead of a swap?
- Leaders flip: gpt-5.6-sol is #1 by tokens (23.3M) but #3 by cost ($33). Is that flip the feature (efficiency story) or disorienting? A $/M-token "unit price" column might tell it better than the toggle.
- Zero rendering: fmtC(0) → "—" but fmtN(0) → "0" (see the "—" model row). Align them?
- Input/output split labels: is "output tokens (incl. reasoning)" the right framing, or should reasoning be its own tile?
- Persistence: URL-only, or should the unit stick across visits (localStorage) like the theme?
Replica of qdash /cost with the c66 unit toggle. Data is a live snapshot (2026-08-14, period=all): by_model/agent/user/team rows, daily series, and /api/stats 24h figures are verbatim; the input/output token split is illustrative (the real split needs the backend totals addition); the two guardrail targets are seeded (live /api/campaigns was empty). Period tabs are cosmetic (all periods show the same snapshot). Row links (agent drill-down, teams focus) and cross-page navigation are stubbed. qdash's local Panel/StatTile/BarList chrome is inlined on layer primitives (DataMetricLabel, UxPercentBar, DataTable, DataChartUI).