On September 3, 2026, OpenAI announced GPT‑6 Astra, with an initial limited rollout and broader access planned over the following days. Looking at the announcement, the benchmarks were not what caught my attention most. It was the change in the question we need to ask inside our companies. Official release notes.

For a long time, interacting with AI was mostly a conversation. You asked. It answered. Someone still had to turn that answer into action.

According to OpenAI, Astra combines browsing, tool use, and software engineering to carry out work across multiple steps, including writing and testing code. That brings AI closer to doing the work that used to follow the conversation. GPT‑6 Astra announcement.

We are moving beyond “what can AI do?” and starting to ask “what are we willing to let AI do?”

Delegation takes more than a good prompt

For me, this is the most relevant part for technology leaders. The more execution capability we give a system, the more explicit we need to be about what we expect from it — and which decisions remain ours.

Asking AI to investigate a production issue may sound clear enough. But does investigation only mean reading logs? Can it change a setting? Restart a service? Notify a customer? Each action has different consequences, even when all of them seem to serve the same goal.

I would start designing that delegation around five questions:

  • Context: what is the goal, and how will we recognize a correct result?
  • Autonomy: which decisions can AI make on its own?
  • Boundaries: which data, systems, and actions are outside its scope?
  • Observability: how do we track what happened and stop an execution?
  • Accountability: who reviews the outcome and takes responsibility for its consequences?

These questions help turn a technical capability into a way of working that the team can understand, follow, and improve.

What the Critical threshold puts in perspective

One detail of the launch makes this discussion particularly tangible. OpenAI says Astra is its first model to reach the Critical level of cybersecurity capability under its Preparedness Framework. This classification concerns evaluated capabilities and risks that require safeguards; it is not a safety certification for every use case. GPT‑6 Astra system card.

The document links this advance to additional controls against malicious use and unauthorized actions by the model itself. It also describes limitations in monitoring. My reading is that capability, authorization, and oversight need to be discussed together.

For a company, that does not mean every task carries the same risk. Preparing a draft, proposing a code change, and making a production change call for different levels of autonomy. The workflow should reflect that difference.

The next decision belongs to leadership

I see room to start with tasks that have a clear scope, verifiable outcomes, and an opportunity for review. As the team learns, it can expand autonomy based on what it observes, while keeping the points that require human judgment explicit.

Leadership now includes designing those decisions: where AI can proceed, when it should stop, and who takes over next.

The next phase of AI will also be defined by our ability to decide how, where, and how far to delegate.

Smarter models expand what is possible. Turning those possibilities into reliable work still takes judgment. For people working in technology and leadership, I think this conversation is only getting started.