Skip to content

Usage

The Usage page reports how many tokens your deployment has consumed and which models spent them. It is a reporting surface, not an enforcement one — it tells you what happened, it does not cap anything. The base view needs analytics:read, which every tier has; the charts and per-model breakdown are a Pro capability, marked below.

All figures are counted in tokens. DioscHub does not price model usage in currency; there are no dollar amounts anywhere on this page.

The Usage page: overview cards (4.54M total tokens, 30 active users), the token-over-time and distribution charts, and the Tokens-by-Model pie with a per-model breakdown.

The Usage page on a deployment with real traffic: overview cards on top, the token-over-time trend, and the per-model breakdown below.

Three controls at the top of the page scope everything below them. Changing any of them re-fetches the report.

FilterOptions
AssistantAll Assistants, or a single Assistant by name.
Date RangeToday, Yesterday, Last 7 Days, Last 30 Days (default), This Month, Last Month.
Group ByDay, Week, or Month — the bucket size for the time-series charts.

The header shows the resolved date range (for example, “Jul 1, 2026 to Jul 31, 2026”) so you can see exactly what window the numbers cover.

Four cards summarise the selected window:

  • Total Tokens — every token spent in the window, with the session count beneath it.
  • Input Tokens — tokens sent to the model, with their share of the total.
  • Output Tokens — tokens generated by the model, with their share of the total.
  • Active Users — distinct users who drove traffic, with the count of active assistants beneath.

The charts and per-model breakdown require a Pro or Enterprise license (detailedUsageReporting). See Licensing & tiers.

Token Usage Over Time is a line chart of total tokens per bucket, with input and output drawn as dashed lines beneath the total. Token Distribution stacks the same input and output tokens as bars per bucket, so you can read the split at a glance.

The model breakdown answers which models spent the tokens. Tokens by Model is a pie chart, sliced per model. Beside it, Model Details lists each model with its token total and its percentage of the whole — useful when a single expensive model dominates the bill.

The Token Breakdown card at the bottom restates input tokens, output tokens, and interactions side by side, with the average interactions per session — a quick read on how much work each conversation costs.

The overview also shows a Failovers count: the number of turns a fallback model answered after the primary model failed. Tokens from such a turn are counted against the model that actually ran. See fallback models.

This page measures consumption after the fact. It does not limit anyone. Capping how many tokens an individual can spend is done with token budgets, which are configured and enforced on the Rate limiting & cost controls page — budgets resolve per Role, per Assistant, or from a platform default, and block the next turn when the cap is hit. Use Usage to see where the tokens went; use rate limiting to bound where they go next.

The assistant used to get a hint on every budgeted turn saying roughly how many response tokens were left. That hint is gone. It gave the size of the next reply, not the quota, and it called a per-window budget a “session”. A status block replaces it, and the block is empty on most turns:

  • Below 80% of the quota, and with no context trigger, the assistant is told nothing about usage.
  • From 80% and from 95%, the turn that crosses the level carries the percent used, the tokens remaining and the reset time, and the assistant says so once. Later turns only ask it to keep answers concise.
  • When the conversation is compacted or close to the model’s context limit, the assistant suggests a new chat, once.

For exact figures the assistant calls its get_my_usage tool when a user asks. The limit on the length of each reply (max_tokens) is unchanged. Details, fields and edge cases are in What the assistant knows about the budget. This changes what the assistant says, not what the Usage page reports.

For the qualitative side of the same traffic — what users asked, and where the assistant fell short — see Conversation intelligence.