See what a run cost, in dollars

Token counts are not a number anyone can reason about, and on your own key they are your money.

The run header now carries an estimated cost next to the model. It prices the tokens that run actually recorded, per model, with the cached share of the input billed at the cached rate. A long run is mostly cached context, so getting that split wrong overstates the figure exactly where the figure matters.

It shows nothing when there is nothing honest to show. A run with no recorded usage, or one on a model with no published price, shows its token count and says why, rather than a blank space that reads as "this run was free".

Prices live on the server and are handed to the console, so the number in front of you and the number the platform computed cannot drift apart between deploys. It stays an estimate at published list prices, and on your own key it is a part of your own OpenAI bill rather than an invoice from us.