GPT-6 Astra can work through tool delays, but the app still runs the jobs
OpenAI's async tool interface separates continued reasoning from responsibility for execution.
OpenAI's documentation describes async tool calling for GPT-6 Astra: a model can continue with independent work while an application runs a slow function or custom tool. Developers enable it with async: true and return the result later using the original call_id.
The distinction matters for teams building agents. OpenAI says this feature neither executes the application's tools nor manages its background jobs. A developer-defined wait tool can let the model pause when a missing result becomes necessary; that helper is not a built-in Responses tool.
The documented compatibility limits also exclude hosted built-in tools from async execution. This is a coordination feature, not a promise that every workflow becomes faster. Applications still need to track pending work and deliver real results before dependent steps proceed.
프리즘코리아 편집국 > Daniel Reed
