Cost Tracking
CubePlex tracks LLM token usage across your organization so you can monitor load and spending. The usage dashboard is available at Admin > Insights (/admin/insights; /admin/cost redirects there).
What gets recorded does not depend on your edition — every deployment tracks usage on every call, and the data is yours in your own database. The dashboard and CSV export that read it back are Enterprise features. Without a licence the Insights page shows an "Enterprise feature" notice; see Editions & Licensing.
What is tracked
Every LLM call made through CubePlex records:
- Input tokens — the number of tokens sent to the model (prompt, system instructions, tool results, etc.).
- Output tokens — the number of tokens the model generated in its response.
- Cache read / cache write tokens — tokens served from or written to the prompt cache, billed at their own rates (often much cheaper than fresh input tokens).
- Provider and model — which provider and model handled the call.
- Timestamps — when the call started and ended.
Each call also records the per-model price rates (input, output, cache read, and cache write) as a snapshot at the time of the call, so historical costs stay accurate even after you later change a model's rates in Model Management.
Cost tracking also records sandbox compute events, so the totals reflect more than just LLM token spend.
Usage dashboard
The dashboard defaults to token metrics so it stays useful even when model prices are not configured (costs would otherwise show as $0). Use the Tokens | Cost toggle in the top bar to switch to USD when you have set pricing under Models. Your last choice is remembered in the browser.
Primary metric
| Mode | KPIs and charts | When to use |
|---|---|---|
| Tokens (default) | Total / input / output tokens, avg tokens per call, stacked series ranked by tokens | Always meaningful; no pricing required |
| Cost | Total cost, avg $/call, series ranked by cost | After model prices are set |
If you switch to Cost while total cost is $0 but tokens are non-zero, the page shows a short hint linking to Model Management.
Spend / usage over time
A time-series chart at daily or weekly granularity, broken down by workspace, model, or user. In tokens mode the axis is token totals; in cost mode it is USD. Use this to spot trends — increasing usage after onboarding, spikes from large batch tasks, or the impact of switching models.
Breakdowns
Alongside the org-wide totals, the dashboard breaks usage down for the selected period:
- By model — which models drive the most traffic (or spend).
- By workspace — which teams or projects are busiest (workspace name, not the internal id).
- By user — per-person usage (display name when set, otherwise email).
- Cache efficiency — hit rate from token fields (token-based in both modes).
Each table row shows input/output (and cache where relevant) token counts, call count, and either a token total or cost depending on the active metric. Workspace and user charts, tables, and filters use the same readable labels.
Summary KPI tiles use org-wide totals for the date range. Workspace/model filters currently apply to timeseries charts. Matching filters on the summary is a known follow-up.
Export
You can export the raw cost data as CSV — for the whole organization or for a single workspace — to analyze it in a spreadsheet or feed it into your own reporting.
Capture: The Admin > Insights dashboard with the Tokens | Cost toggle, token KPIs (or cost KPIs), stacked chart, and breakdown table.
Asset: /img/admin/cost-dashboard.png
Tips for managing costs
- Set accurate cost rates. Make sure the input and output token rates in Model Management reflect your actual provider pricing. Without rates, cost mode shows $0; tokens mode still reflects real load.
- Review regularly. Check the dashboard weekly to catch unexpected spikes early.
- Right-size your models. If a significant portion of usage comes from a high-cost model on simple tasks, guide the team toward a lighter model for those cases.
- Use the per-model view to compare. After switching models or adding a new one, the per-model breakdown shows whether the change had the expected impact.