The policy is the same for every outcome, because billing is metered on actual work: you’re charged for the AI work that actually happened, and only that.

​
If you cancel a run

Hit Stop at any time during a generation. The run winds down and you’re billed only for the work completed before you stopped — never the full projected build. The chat confirms it: “Generation cancelled — you were only charged for the work completed before you stopped.” Details on how cancelling works: Canceling generations.

​
If a run fails

Sometimes a generation errors out — the AI hits a wall, a build step crashes, or an upstream model provider has an outage. When that happens:
  • You’re billed for the tokens actually consumed up to the failure, at the normal rates. Work the AI completed before failing (assets generated, code written) is usually still in your project.
  • You are not billed for anything after the failure point, and failed runs don’t incur any penalty or fee.
Why not fully refund failed runs? Because “failed” usually isn’t all-or-nothing — a run that errors on step 9 of 10 still did 90% of the work, and that work is in your game. Automatic full refunds would also make it possible to get unlimited free AI work by forcing errors. Metering actual usage keeps pricing honest in both directions.

​
If a run stops to ask you something

When the AI asks a clarifying question or presents a plan for approval, the session pauses. Work done up to the pause is billed; nothing is metered while it waits for you.

​
If the platform is at fault

If a generation charged you but produced nothing — the engine crashed before doing any work, an infrastructure error ate the session, the game never updated — that’s on us. Contact support with the game name and rough time of the prompt (Getting help), and the team can review the session transcript and make it right with a credit adjustment.
Before writing in, check your Usage page: expanding the charge shows exactly how many tokens were processed. A charge with millions of tokens means the AI really did work (even if the result missed the mark — that’s a prompting problem we can help with); a charge on a session that visibly did nothing is worth flagging.