Claude account chats stop wasting calls and count what they spend
Claude account steps no longer pay for hidden extras, and summaries, side calls and cut-off steps now show up in your usage instead of as zeros.
Every step of a Claude account chat starts its own Claude Code process. Some Claude Code versions used that moment to make a small extra request of their own, just to name the terminal window, and it was paid from your plan. Bandicot now switches that off for every step, and turns off Claude Code's other non-essential traffic for the steps that answer you.
Some of that spend was also invisible. When a conversation grows long, Bandicot writes a summary with a model, and that call was never counted. It now appears in the chat's usage, tagged as a summary, and the efficiency view counts how many summaries ran and what they used. It is not counted as a turn, and the model and effort it uses are unchanged.
Claude Code also reports how much its side requests used. Bandicot now shows that as its own number next to a step's tokens instead of folding it in, so your step totals still mean what the step used.
Steps that were cut short
A step that was cut off before Claude Code finished it used to be recorded as zero tokens, even though it had been paid for. These now record what the stream showed. When the model had not yet reported its final output count, Bandicot records a conservative floor from the text it saw.
Bandicot also ends a step sooner when the model keeps writing long prose after its tool calls, which nobody reads and which still costs tokens. Short notes between calls are kept, and so are all the calls.
The usage panel stops asking for fresh numbers while the window is hidden, and catches up once the moment you look at it again.