Usage

The Usage view shows resource consumption for your organization — LLM tokens, compute hours, storage, and platform activity (threads created, messages sent). It is read-only and reflects the state at page load.

Open Usage

Console → Activity → Usage (/organizations/<org>/activity/usage).

The view is a single scrollable page with four sections — LLM, Compute, Storage, Platform — and a time range selector at the top.

Time range

RangeBucket size
Last 24h5-minute buckets
Last 7d1-hour buckets
Last 30d6-hour buckets
CustomBucket auto-picked to keep charts legible

Usage is not live — to refresh, change the time range or reload the page.

LLM

The most-watched section. Summary cards:

  • Input tokens — sum of prompt tokens across all LLM calls.
  • Cached tokens — <cached> of <input>. Provider-side prompt cache reuse, when supported by the model.
  • Output tokens — what the model generated.
  • Successful requests / Failed requests — counts.

Charts:

  • Tokens over time — stacked bars: cached + fresh input vs. output, per bucket.
  • Top consumers — horizontal bars by consuming identity (which agent or user drove the calls).
  • By model — horizontal bars by model.

Cache efficiency is a leading indicator of cost — high cached-token ratios mean you are not paying full price for the prompt every call.

Compute

Time agent workloads spent occupying a flavor, in flavor-hours.

  • Summary cards — flavor-hours over the selected range.
  • Usage over time — bars per bucket, broken down by flavor.
  • Top agents — horizontal bars by agent.

Compute is billed by which flavor a workload occupied and for how long, not by what it actually used. A flavor is recorded against the runner that declared it, because the same name on two runners need not describe the same resources. The flavor is fixed when the workload starts: repointing its environment afterwards does not change what a running workload bills.

Storage

Persistent volume storage in GB-hours.

  • Summary card — Storage-GB-hours.
  • Usage over time — bar per bucket.
  • Top agents — horizontal bars by agent.

Like compute, storage is allocation-based — the size of the provisioned PVC times its lifetime.

Platform activity

Threads and messages — the lightest section, useful for capacity planning.

  • Summary cards — Threads created, Messages sent.
  • Activity over time — side-by-side bars.

What's missing

The Usage view focuses on aggregates. For per-run inspection, use the Run Timeline. For provider-side billing, use your provider's invoice — the platform's token counts are accurate but the dollar figures depend on your provider's pricing.

Cluster admins

Cluster admins see one Usage view per organization (Console → org → Usage). There is no cross-org aggregate in the Console; query the Metering service directly for that.