Reduce the tool bottleneck that used to pause a turn
Traditional function calling normally pauses the model's turn until a tool result arrives. With GPT‑6 Astra, marking a function or custom tool `async: true` lets the model answer an independent part, call another tool or prepare later work while the application runs a slow lookup or job.
Execution does not move to OpenAI. The application still owns queues, failures and retries, then returns the completed output in a later Responses request using the original `call_id`. This is also distinct from Background mode, which makes response generation asynchronous.
Change direction before the work finishes
Mid-turn steering lets a user add a requirement or redirect an Astra response over WebSocket. During a long research run, the user might say, “Exclude Europe and prioritize small-business cases.” Completed work is preserved and the update is included in a continuation.
Steering does not retract output already delivered, undo an action or cancel a tool that has started. An accepted event only means the update was queued, not that the model has applied it. The application must define cancellation and decide whether a late result still belongs to the latest task version.
From sequential assistant to interactive worker
Together, the features support long research, multiple data lookups, drafting and review while a user adjusts priorities from intermediate results. They also require a state machine for pending calls, late outputs, duplicate continuations and spending limits.
A strong interface should expose each job's state, original call ID, instruction version, cancellation status and whether its output was accepted. The innovation is not merely hiding latency; it is making ongoing work visible and steerable.