Sending the same kind of prompt through Auto, a developer switches the optimisation mode from Cost to Intelligence and watches the per-request charge rise sharply. What accounts for the difference?

The principle — Auto is not a single pricing scheme. The cost-optimised mode keeps the bundled Auto rate, charged the same way whichever model handles the request. The higher optimisation modes instead bill at the rate of the model the router actually chose, so the price follows the routing.

Why the key is correct — Once the mode is raised, the router is free to send work to more capable models, and the charge for each request follows that model's own rate rather than a bundled figure. The prompt did not change and the volume of work did not change; the billing basis did. That is why the same task can cost several times more after a mode switch, and why an included budget can drain far earlier in the cycle than the request count would suggest.

Why the others are wrong — Believing Auto carries one flat rate in every mode sends the developer hunting through prompt length and context size, neither of which explains the jump. A fixed per-session premium would be a small, predictable addition, whereas this effect compounds across every routed request. Changing the mode also does not move an account onto on-demand billing: included usage is still drawn down first, and charges above the subscription begin only when that budget is exhausted, whatever mode is selected.

Remember this — The mode decides the meter. The cost mode keeps the bundle; the higher modes pay the routed model's price.

Sources — Cursor's usage-based charges documentation.