Skip to content
2 minute readNewAgent-ready

The plan quote now bills a rented machine by the clock, waiting included

A step in the plan quote can now be a machine rented by the hour, such as a GPU. The quote bills boot, work, the minutes spent waiting for answers and the billing step, says what waiting cost, and gives the longest wait worth keeping the machine up for.

The plan page's "Where the rented minutes go" bar for a rented GPU, with 60 minutes of work, 10 of boot and 10 of waiting, and the cost of waiting

A rented GPU bills the clock, not the work. It bills while it boots, while the agent waits for someone to answer a question, and up to the next billing step. A machine that stops itself after a few idle minutes has to boot again when the answer finally comes. The plan quote on /plan-cost now prices all of it.

  • Rented by the hour. Set a step to per hour, with its hourly rate and the minutes one unit of work takes. Add the boot minutes, when the machine stops itself, the billing step (second, minute or hour) and each wait for an answer.
  • Where the minutes go. A new bar splits the billed minutes into work, boot, waiting and rounding, and says what waiting cost, new boots included.
  • When to stop the machine. The quote gives the longest wait worth keeping the machine up for. A longer wait costs less with the machine stopped. Asking before renting shows up in the savings.
  • Per generation or per hour. Another way to do a step can now be charged differently, so paying per generation can be compared with renting the machine.
  • Agents can ask too. POST /api/public/plan-cost and the tool estimate_plan_cost take the same inputs, listed in the [API docs](/api-docs).

Works with the MCP tool estimate_plan_cost and the @last-price/tools-* packages.