Skip to main content
/admin/cost is a single-page operational view of how much your AI providers are costing you. It reads the ai_call table (every call your app makes is logged there with token counts, cost in micro-cents, and latency), rolls up the last 30 days, and breaks it down per provider/model. Auth-gated by requireAdmin() in the layout — only admins reach the route.

​
Prerequisites

  • AI calls already flowing through the provider abstraction (Phase 2). The page is empty until your first call lands.
  • Your user record has role = 'admin'. Set it via SQL or the admin user page.
  • The ai_call table exists — created by pnpm db:push after src/db/ai.schema.ts.
  • Per-model price book entries (otherwise cost shows as 0 — see pitfall 3).

​
Step-by-step

  1. Sign in as an admin at /login, then visit /admin/cost.
  2. Read the four summary cards at the top:
    • Total calls (last 30d)
    • Total tokens (input + output)
    • Total spend (rendered through <CostBadge>, micro-cents to dollars)
    • Success rate (% of calls with status = 'success')
  3. Drill into the per-model table below. Rows are sorted by spend descending. Columns: provider/model, calls, input tokens, output tokens, cost (CostBadge), avg TTFT (ms), error rate (%).
  4. Spot the offenders — high error rate, slow TTFT, or runaway spend.
  5. Investigate in /admin/orders or by querying ai_call directly with pnpm db:studio.

​
How the rollup works

The page runs two SQL aggregates against ai_call: 
// Summary card totals
const totals = await db
  .select({
    calls: sql<number>`count(*)::int`,
    successCalls: sql<number>`count(*) filter (where status = 'success')::int`,
    inputTokens: sql<number>`coalesce(sum(input_tokens), 0)::int`,
    outputTokens: sql<number>`coalesce(sum(output_tokens), 0)::int`,
    cost: sql<number>`coalesce(sum(cost_micro_cents), 0)::int`,
  })
  .from(aiCall)
  .where(gte(aiCall.createdAt, since));
The per-model table groups by (provider, model) and orders by sum(cost_micro_cents) desc. Both queries window on createdAt >= now() - 30d.

​
Verify it works

  1. Trigger an AI call from anywhere in the app (e.g. a test prompt).
  2. Refresh /admin/cost — the call count goes up by one, tokens reflect the request/response sizes, and cost is non-zero (assuming the model is in your price book).
  3. Confirm the model appears in the per-model table with the right provider.
  4. If you have multiple providers configured, run a few calls against each; the rows sort by spend so the most expensive model bubbles up.

​
Common pitfalls

  1. Empty dashboard until you make AI calls. New install? Of course it’s empty. The page only reads from ai_call — no calls, no rows.
  2. Cost shown as 0. Your cost_micro_cents column is 0 because the model has no entry in the price book. Add it under src/ai/pricing.ts (Phase 2) — or accept the discrepancy if you only care about token counts.
  3. Micro-cents math. Stored cost is in micro-cents to avoid floating- point errors. <CostBadge> divides by 1,000,000 to render dollars. If you query the table directly, remember to do the same: cost_micro_cents / 1_000_000.0 for USD.
  4. Slow queries on large datasets. Past ~1M ai_call rows the per-model group-by gets sluggish. Add a Postgres index on (created_at, provider, model) or partition by month.
  5. 30-day window is hardcoded. Want a different window? Edit WINDOW_MS at the top of src/app/[locale]/(app)/admin/cost/page.tsx. There’s no query-string toggle yet.

​
Official docs