Skip to content

Cost and usage ​

Humanize counts the tokens every agent uses and prices them in US dollars. Watch these figures while a chat or an epic runs, or add them up afterward by day and by flow.

In Humanize Desktop, the Usage ring under a chat's box is clicked, a popup shows what the chat and the folder spent, and Usage settings opens the folder's Usage pageIn Humanize Desktop, the Usage ring under a chat's box is clicked, a popup shows what the chat and the folder spent, and Usage settings opens the folder's Usage page

What is counted ​

Humanize reads spending from the coding agent's log during each turn and updates the figures at least every five seconds.

FigureWhat it is
TokensCounted per kind: input, output, cache read, cache write and reasoning.
CostUS dollars at list price, per model. See Prices.
RateOutput tokens per second, per model, over the last five minutes.
TimeHow long agents worked.

Minimum figures show + after them in the terminal and Usage popup, and ≥ before them in the app's Spending table. A token count is a minimum when a coding agent does not report that kind. A cost is a minimum when some models have a price and others have none.

In the app ​

Open the Usage ring ​

Click the ring at the end of the row under a chat's box. The popup shows:

  • This chat: Spent, Tokens, Time, and each token kind with its share while the chat runs. Once the chat rests, it shows Spent, Output tokens, and Time working.
  • The folder: what all chats and epics in it spent Today (UTC) and over the Last 7 days.
  • Usage settings, which opens that folder's Usage page, in its Workspace settings.

The Usage popup: what this chat spent, then what the folder shop spent today and over the last 7 days, and the Usage settings linkThe Usage popup: what this chat spent, then what the folder shop spent today and over the last 7 days, and the Usage settings link

On the New chat page, the ring shows only the folder's figures. With no folder picked, and in a chat that has no folder, the folder part reads Every folder and Usage settings opens the daemon's Usage page.

Watch a run ​

Open an epic and check the Epic tab in the side panel. Figures shows time, spent, tokens, and the budget. Spending lists one row per model with tokens of each kind, Cost, and Out/s, followed by an All row. Once the epic ends, Figures shows what it spent, its output tokens, and how long it took. In a chat, the Details pane shows the same under Usage. See Budgets for budget meters and Watch a run for the rest of the panel.

The Spending table of a running epic: Opus 4.1 has tokens and a rate but a dash for cost, Haiku 4.5 has a costThe Spending table of a running epic: Opus 4.1 has tokens and a rate but a dash for cost, Haiku 4.5 has a cost

A model with no price shows a dash, not $0.

View the Usage pages ​

There are two Usage pages with the same layout.

PageWhat it countsOpens from
Usage in Settings, under DaemonEvery chat and epic the daemon shown ran, across all folders or none.Usage in the daemon bar menu at the foot of the sidebar, or Usage settings in the Usage popup when no folder is selected.
Usage in a folder's Workspace settingsThat one folder, named at the top of the dialog.Usage in a folder's menu, Workspace settings in its Group and sort menu, or Usage settings in the Usage popup of a chat in a folder.

The Usage page in the Workspace settings of the folder shop in Humanize Desktop: the folder named at the top, the Range buttons, tiles for spent, output tokens, agent time and ended, and a bar chart of cost by dayThe Usage page in the Workspace settings of the folder shop in Humanize Desktop: the folder named at the top, the Range buttons, tiles for spent, output tokens, agent time and ended, and a bar chart of cost by day

PartWhat it shows
RangeThe lookback window: 7d, 30d, 90d or 365d (default 30 days). This is the only control on the page.
TilesSpent and epic count, Output tokens, Agent time, and Ended (runs grouped by outcome).
Cost by dayOne bar per UTC day. Open Table below the chart to see epics, spend and output per day.
By flowOne row per flow, ordered by cost: epics, spend, share, output and time. Chats appear under the flow chat. Click a flow name to open Search on its epics, or click chat to open the Chats tab in Search. From the Workspace page, Search is filtered to that folder.

The By flow table: ralph_loop, rlar, chat and goal, each with epics, spent, a share bar, output and timeThe By flow table: ralph_loop, rlar, chat and goal, each with epics, spent, a share bar, output and time

Both pages read chat and epic records directly from the daemon and refresh every 30 seconds. When a range has no runs, the page shows Nothing ran in this range. with a New chat button.

In the terminal ​

Watch a run ​

Once an agent spends tokens, two lines appear above the editor and under the agent lines: tokens by kind, then cost and rate.

The terminal interface during an epic: above the editor, a line of tokens per kind and a line with the cost and the output rateThe terminal interface during an epic: above the editor, a line of tokens per kind and a line with the cost and the output rate

Press left to open the monitor. Under the graph, Tokens lists each model with its tokens, cost, and rate. Kinds lists the tokens of each kind for the whole epic.

The monitor of a running epic: under the two agents, Tokens lists two models with tokens, dollars and out/s, and Kinds lists input, output, cache_read and cache_writeThe monitor of a running epic: under the two agents, Tokens lists two models with tokens, dollars and out/s, and Kinds lists input, output, cache_read and cache_write

View the usage page ​

Type /settings usage.

The terminal's usage page: Range and Workspace rows, the total spent, one bar per day and one row per flowThe terminal's usage page: Range and Workspace rows, the total spent, one bar per day and one row per flow

RowChoices
Range7 days, 30 days, 90 days or 365 days. Opens on 30 days.
Workspacethis directory or every directory. Opens on this directory.

Below the options, you see the total (cost, epics, output tokens, time worked, and how they ended), each UTC day with an epic (newest first), and each flow with its share (costliest first). This page counts epics only, not chats, and has no save button because it stores no settings.

From the command line ​

Run hmz exec or hmz chat to see a summary line under each turn: tokens by kind, cost (when priced), model, and agent.

hmz exec running the rlar flow: under each turn a line reads input, output, a dollar cost, the model and the agenthmz exec running the rlar flow: under each turn a line reads input, output, a dollar cost, the model and the agent

Pass --json to hmz exec to get token counts in the result event of each turn: tokens (by model) and spent (by kind), with no dollar figure. See Events and records.

Prices ​

Humanize calculates cost at list price from the OpenLLMPrices list (openllmprices.com). It fetches the price list in the background, caches it locally, and refreshes it after 24 hours.

HUMANIZE_PRICESEffect
not setFetch the list from OpenLLMPrices.
a URL or a file pathRead that list instead.
off, 0, no, none or emptyFetch nothing.

An unpriced model shows tokens and no money, never $0.00. It counts as zero against a cost limit; see Budgets for how Humanize reports this.

Good to know ​

  • Cost is a list-price estimate per token, even on subscription plans, not your provider's bill.
  • Reasoning tokens are priced as output. A cache write with no price of its own is priced as input.
  • Cost follows the list's dearer rates where they apply. A write to a one-hour cache is priced at the list's one-hour price, and a request whose prompt (input, cached or not) is past a model's long-context line, such as more than 272k tokens, is priced at that tier's rates.
  • Unpriced cache reads are excluded, making the total cost a minimum.
  • A model is matched to the price list by name, ignoring a release date, -latest, and a cloud prefix such as bedrock-.
  • Money formats as $0.0042 under a cent, $1.56 from a cent, and $1,234 from $100. Tokens format as 999, 12.3k, and 1.25M. Durations format as 42s, 3m07s, and 2h05m.
  • Every day is a UTC day.
  • A run killed before it records its spending counts as a run with nothing spent.

Open source on github.com/humanfia · RSS: blog · news