Appearance

Every time an agent thinks, it makes a call to a language model that costs money and takes time. Frona records each of those calls so you can see exactly where your tokens and dollars go — live in the chat header, and broken down in a dedicated dashboard.

This is per-user: you only ever see your own usage. There's no shared or server-wide rollup.

The usage pill ​

Once an agent has made at least one model call in a conversation, a small pill appears in the chat header next to the agent's name. It updates live as the conversation continues, and shows:

  • ↑ input tokens (with the cached portion in parentheses)
  • ↓ output tokens
  • cost for the conversation so far
  • context-window saturation — how full the agent's context is, as used / window (%)

Hover over the pill for a breakdown: a context bar that turns yellow as it fills, the split between cached and fresh input, total output, cost, and the number of model calls and tool calls in the conversation. If the last call fell back to a backup model, a ⚠ appears.

The context-window figure is a useful early warning. When it climbs toward the limit, the platform summarizes older messages to make room — the pill is where you can watch that happen.

The usage dashboard ​

For the bigger picture across all your conversations, open Usage from your profile menu. A time-range selector (24 hours, 7 days, 30 days, or 90 days — default 30) sits at the top and applies to every tab.

Overview ​

A high-level summary: total cost, number of API requests, tokens in and out (with the percentage served from cache), and a p95 duration. Below that, tokens over time, a breakdown of cost by call type, and your most expensive conversations in the window.

Cost ​

Where your spend goes — total spend, average cost per request, blended cost per million tokens, and request count. Charts break spend down over time, by call type, and by model, with a full list of your priciest conversations.

Latency ​

Wall-time percentiles (p50 / p95 / p99) for two things:

  • Total call duration — request sent to final byte, including any retries and backoff.
  • Time to first token — how long until the agent starts streaming a reply. This is the "feels responsive" number. Background calls like title generation don't count toward it.

A per-model table compares both across the models you use.

Tokens ​

Input, output, and cached tokens over time and by model, plus a cache-hit ratio chart. Cached input is far cheaper than fresh input, so a higher cache-hit ratio means a lower bill.

What counts as a call ​

Not every model call is a reply to you. The dashboard groups calls by kind so you can see what's driving cost:

  • Replies and tool turns — the agent answering you or working through tools.
  • Title generation — naming a new chat.
  • Compaction — summarizing older messages to free up context.
  • Signals and routing — background bookkeeping.

Each recorded call notes which model handled it and whether a fallback model was used, so an unexpected spend spike usually traces back to a specific model or a specific conversation.

Where cost numbers come from ​

Cost is calculated from a pricing snapshot taken when each call is recorded, using model pricing from the model catalog. If pricing for a model isn't known, its calls still show token and latency figures — they just don't contribute to the cost totals.

Next steps ​

Frona AI