48.4k Log in Get started
Dashboard

What goes in an LLM spend overview dashboard?

An LLM spend overview dashboard is the page leadership actually opens: what the LLM layer costs, where the spend concentrates, and whether growth is usage or waste. It's built from daily usage rollups synced from a gateway or tracing tool — not from invoices, which arrive too late to act on.

For: Engineering leadership, finance partners, and platform teams. Refresh: daily. Source: modeled LLM usage tables in a Metabase-connected database.

Which cards belong on this dashboard?

  • Total LLM spend, trailing 30 days (number + trend)
  • Spend per day and per week (line)
  • Spend by model (bar)
  • Spend by app or team (stacked bar)
  • Month-over-month spend growth (number + trend)
  • Cost per request trend (line)
  • Spend on failed or retried requests (bar)
  • Spend vs. budget by team, where budgets exist (table)

What data does this dashboard need?

  • Daily usage rollups per model and app: requests, prompt and completion tokens, cost
  • App, feature, and team tags carried from request metadata
  • Error and retry counts so failed-request spend is visible
  • Budgets or targets per team for the variance view
  • Cache and batch discount fields where the source records them

How do you build it?

  1. Sync usage rollups from your LLM gateway, tracing, or eval tool into a database, keeping stable IDs, timestamps, token counts, and costs.
  2. Build clean models at the grain this dashboard requires — daily per app and model is the workhorse.
  3. Create one saved question per card and certify the shared models and metric definitions.
  4. Add dashboard filters for Date range, model, app or team, environment, provider.
  5. Exclude test, staging, and playground traffic from headline cards, and show the refresh time.

Example card SQL

Daily LLM spend by model with month-over-month growthPostgreSQL
WITH daily AS (
  SELECT
    date_trunc('day', window_start) AS day,
    model,
    SUM(cost_usd) AS cost_usd
  FROM llm_request_rollups
  WHERE environment = 'production'
  GROUP BY 1, model
)
SELECT
  date_trunc('month', day) AS month,
  model,
  ROUND(SUM(cost_usd), 2) AS cost_usd,
  ROUND(
    100.0 * (SUM(cost_usd) - LAG(SUM(cost_usd)) OVER (
      PARTITION BY model ORDER BY date_trunc('month', day)
    )) / NULLIF(LAG(SUM(cost_usd)) OVER (
      PARTITION BY model ORDER BY date_trunc('month', day)
    ), 0), 1
  ) AS mom_growth_pct
FROM daily
GROUP BY 1, model
ORDER BY 1, cost_usd DESC;

Dashboards

Model usage & cost dashboardLLM latency & errors dashboardLLM adoption & unit economics dashboardEval quality tracking dashboard

Integrations

AI & LLM analytics integrationsLangfuse + MetabaseLangSmith + MetabaseHelicone + MetabaseBraintrust + MetabaseArize Phoenix + MetabaseW&B Weave + MetabaseOpenRouter + Metabase

Metrics

Token usageCost per 1M tokensLLM cost per userp95 LLM latencyLLM error rateCache hit rateEval pass rateLLM requests per day

Analytics

LLM analytics

FAQ

Should the dashboard read from invoices or request logs?
Request logs. Gateways and tracing tools such as Helicone, Langfuse, and OpenRouter price each call at request time, which gives you daily, per-model, per-app signal you can act on. Invoices lag by weeks and lose the attribution. Reconcile the two monthly and treat small gaps — caching discounts, credits, minimums — as expected.
How do I attribute spend to teams and features?
Tag requests at the source: most proxies and tracing SDKs accept custom metadata (app, feature, team) on every call. Carry those tags through the rollups and attribution becomes a GROUP BY. Retrofitting attribution onto untagged history is guesswork — start tagging before you need the chart.
Is rising LLM spend automatically a problem?
No — spend that grows slower than usage is efficiency. That's why this dashboard pairs total spend with cost per 1M tokens and cost per request: flat unit costs with rising volume is healthy growth; rising unit costs point at prompt bloat, expensive-model drift, or a broken cache.
What data window should headline cards use?
Trailing 30 days for the headline, with a same-length prior period for the delta. Calendar months make the number jump around with month length, and some sources — OpenRouter's activity API, for example — only retain a 30-day window, so a scheduled daily export is what builds anything longer.