Usage and cost

Session details shows this chat, the Efficiency pane this session's story, AI Gateway › Usage your account limits and spend, and AI Gateway › Activity and Habits your history.

Everything here is read from this Mac. Subscription use has no price, so for a signed-in account the figures are tokens and limits, not money.

Open Session details beside the chat; its Usage section adds up the turns, model calls, tokens and cost of this session.

A bar chart shows tokens per turn; hover a bar for that turn's tokens, tools and cost. A cost marked Estimated means a provider did not report its price, so that figure is partly estimated.

Hover a chat that started sub-sessions to see what the whole tree spent, and how much of it was sub-sessions.

The chip in the prompt bar shows how much of the model's context window this chat fills; press it to see more.

Session details › Context breaks it down: system prompt, tool definitions, subagents, summaries, messages and free space. Compact session there shrinks the chat now; Bandicot also compacts by itself at a threshold. Sessions explains compaction.

Window unknown means the provider does not say how large the model's window is.

AI Gateway › Usage shows each connected account's limits as a meter with the share left and when it resets.

Meters refresh about once a minute, and Refresh limits reads them now. A meter warns as it fills, and a pace line says when it will run out at the current rate.

With two ChatGPT or two Claude sign-ins, a turn that reaches one account's limit carries on with the other, and a notice says Continuing on the next account.

It only moves between accounts of the same kind. The limited account rests until the time it reported, or a quarter of an hour if it gave none, then comes back. A turn moves a few times at most before it stops with the provider's message.

The Usage tab marks a resting account Limited · back in … and names the account turns continue on.

1. 1Open AI Gateway › Providers.
2. 2On the ChatGPT or Claude card, press Add another ChatGPT account or Add another Claude account.
3. 3Sign in with the second account.

Settings › AI usage holds one allowance, in US dollars at each model's API price, for the AI work Bandicot does on its own.

Automatic work is what you did not type: System One decisions, setup, grading, experiments and reviews. On a signed-in account it only uses that account's own limits; real money is spent only on an API key.

The allowance starts at five dollars and does not reset on its own: raising it adds headroom. Each call reserves its worst case first and settles to the real cost. Turning it off stops jobs already running. The readout shows what is spent and reserved.

Settings has the full tab: AI usage settings.

1. 1Press ⌘, and choose AI usage.
2. 2Type an amount in Automatic AI allowance · USD.
3. 3Leave Enable automatic AI work within this allowance on, or turn it off to stop automatic work.
4. 4Press Save account limit.

Efficiency (⌘8) has three views in this order, Overview, Improve and Timeline, and a status word at the top such as Watching or No activity yet.

Overview leads with what the session spent. Press the cost to see how it is counted: each model call's price, once, including work the session started. Quality, time, tokens, context and System One decisions each open a short explanation of how they are counted. Where the work goes lists the model calls, and Skills and tools shows how often each ran and whether it worked.

These are estimates, not invoices. A line says when some calls were not recorded and the totals are partial, and savings are shown only after a side-by-side test. Hold numbers freezes the figures while you read; Back to live resumes.

The pane covers this session and the work it started. With no session open it reads Open a session to see its numbers.

Improve writes checks for your task, tests a change side by side against the current setup, and applies it only when you approve.

The tab is titled Improve this session. A word at its top shows where it stands: Sign-in needed, Off, Watching, Paused or Allowance used up. Under Automatic AI work, a bar shows the shared allowance from Settings › AI usage and what this session has spent from it. Session settings sets how long evidence is kept and the longest background job.

Pause optimization and Resume optimization stop and restart the tests, Stop the running test cancels one in flight, and Keep improving after the agent answers lets tests continue once a turn ends. Your own work always goes first.

Under What good looks like, Set up checks (later Rewrite the checks) has the session's model write the goal and the checks, and Grade the work so far scores it. Under Proposed change, Run a side-by-side test compares a change with the current setup on cost, time and tokens. Review and apply applies a change that passed; Undo this change goes back.

Every step that spends money asks first: the first press shows the cost and where it is paid from, and the second press runs it. A change that fails a required check never qualifies.

1. 1Turn on automatic AI work and set an allowance in Settings › AI usage.
2. 2Open Efficiency with ⌘8 and choose Improve.
3. 3Press Set up checks, read the cost, and press it again to confirm.
4. 4Open Test cases and add at least one case to learn from and one held back, each with an answer you checked yourself.
5. 5Press Run a side-by-side test, then Review and apply if it qualifies.

Timeline, titled Everything this session did, lists calls, decisions, checks and changes, newest first, with the record behind each.

A count at the top reads, for example, 3 updates. Each row shows the time, an update number and what happened: a model call started or finished, a tool, a skill loaded or used, System One decisions, context trimmed, or an improvement step. Work done by a sub-task is marked with sub-task and its id.

Press a row to open its record and a fingerprint (a SHA-256 of the record, to show it was not changed). Show earlier reads older rows. If new rows arrive while you read older ones, Show latest returns to the live page.

On AI Gateway › Usage, the Usage board shows tokens, cost, turns and tool calls for one profile or all of them, over seven, thirty or ninety days.

The Profile menu starts on your current profile and has All profiles; with other profiles on this Mac it lists them too. Range offers 7 days, 30 days and 90 days. Four tiles show Tokens, Cost, Turns and Tool calls; the cost tile says every turn priced or some turns unpriced. Tokens per day charts input, output and cache reads. With All profiles chosen, a Profiles card ranks them.

Below are cards that rank what you used: Providers, Models, Effort, Work profiles, Agents, System One, Tools, Browser, Skills, MCP servers, Plugins and Jev commands. A card shows its top rows; Show all and Show fewer expand and fold it. A card with nothing in the range is left out and named in one line: Nothing recorded in this range for ….

In Settings › Profiles, each profile also shows its last thirty days as one line, or No use in the last 30 days.

Lower on AI Gateway › Usage, Routing draws accounts, the gateway and the profile that spends, and Agent passes lists each finished agent pass.

The Routing diagram shows your connected accounts feeding the gateway and the profile that spends. An account with a stored credential counts as connected before it has run a prompt.

Agent passes is a local record of passes that finished. Today counts Completed, Interrupted and Failed. Pass history filters by range (Today, Last 7 days, Last 30 days or All recent), by outcome and by workspace, and Refresh reads the ledger again. Interrupted and failed passes stay listed, so the history stays honest.

AI Gateway › Activity shows your use day by day over a year and how each model behaves, and Habits shows when and how you work; both are read from this Mac only.

Activity has a year of days as a grid, then How each model behaves, with a range of 30 days, 90 days or 1 year. Each panel compares models on one habit, and none of them is a score. A bar appears only after thirty samples of that habit; before that the panel says how many it has. With no turns yet it reads No turns in this range yet., and with an older engine Model habits appear after the next engine update.

Habits, titled How you work, reads the active profile's ledger over a range you choose. It shows sessions, turns, tokens, spend and the peak hour, then traits (press one to see the number behind it), tokens by hour, turns by weekday, how sessions end, which tools fill your context, tokens by project folder, a footprint per session and speed to first token. None of it reaches a model.

The profile switch and the cards for providers, models, effort, agents, tools and the rest are on the Usage tab, not here: see Read spend by profile.

On an Anthropic or OpenAI API key, the Slow lane for automations switch runs automation work at half price, and it can take hours.

The switch shows on the provider's card in AI Gateway only for a connected key on the provider's own host. It affects automations, not the chats you type in.