Usage and pricing
One balance, three companies billing against it, and a rule about what we will not guess.
A run spends with three different companies on three different bases. A model charges per token. A hosted browser charges per second it is open — and only while it is open, so a run parked waiting for you is not billed for thinking time. A messaging provider charges per message sent. The Usage page keeps them apart, because one merged figure hides which of the three is the expensive one.
Where no rate has been configured, the figure reads unpriced. That is deliberate and it is not the same as zero. A model call with no rate in the card still cost somebody money; we just cannot say how much, and a confident $0.00 is a lie that compounds over a month.
The same rule applies to a run whose provider cost never came back. The whole amount set aside for it is counted rather than guessed downwards, and the run page says that is what happened.
Each workspace has an amount it may spend. Starting a run sets some of it aside; finishing gives back whatever was not used. A run cannot start if the workspace cannot cover its reservation, which is the point: the ceiling is enforced before the spending, not discovered after it.
Not live yet. There is no payment processor in this product. No card is stored anywhere, there is no checkout and there is no subscription. The pricing page shows planned numbers, marked as placeholders in the code, and every button on it opens a waitlist that takes an email address and nothing else. How an agent would pay for something it buys for you has been decided and not built: a card minted for that one purchase, with the amount you approved as its ceiling, so a bug cannot cost more than the number you read.