OpenAI Agents SDK Stops Billing After Guardrails Reject

OpenAI shipped Agents SDK 0.19.1 on July 29, closing a gap where a rejected guardrail check didn't actually stop the model from running, per the release notes.

That gap meant teams could get billed for a full model turn even after an input guardrail had already flagged the request as bad. The turn kept executing anyway.

The patch cancels the parallel input-guardrail task the moment a model turn fails, and cancels the streamed model call the moment an input guardrail fails, instead of letting both run to completion independently. It also cancels sibling enablement checks on failure, so one failed check stops related checks from finishing pointlessly.

Separately, the same release redacts Realtime audio format diagnostics and Blaxel unmount paths out of the logs, trimming detail that had been leaking into debug output.

This reads like a cost fix wearing a correctness label. Anyone running guardrails on Realtime or streamed agents was probably overpaying every time a rejection fired, and OpenAI's changelog doesn't call out the billing angle, it files the whole cluster under cancellation and cleanup. Worth watching whether future SDK releases keep doing that instead of naming the cost fix directly.