GPT-6 Astra Gives OpenAI Its First Model Built to Run Systems Unsupervised

OpenAI launched GPT-6 Astra this week, calling it its most capable model yet for business use, with advanced reasoning, computer-use abilities and stronger coding and cybersecurity performance built for enterprise workloads.

The company positioned the model as a step past chat-based assistance toward systems that act directly inside software. It published a case study showing one customer already leaning on that capability in live operations, and the clearest signal of what computer-use is meant to do in practice comes from that account.

Perplexity told OpenAI it now uses the model to draft communications, push software changes and monitor production systems, checking its work far less often than it did with prior generations, according to OpenAI.

Rather than answering questions, the system operates within a company’s actual infrastructure, taking actions that previously required a human to execute. It then reports back rather than waiting for step-by-step approval, a meaningful shift in how much autonomy an enterprise is expected to extend to a deployed model.

Why Perplexity’s Trust Signal Stands Out

Perplexity is not a passive customer here. It builds its own AI-powered search and answer products, competing in the same broader market OpenAI serves through ChatGPT, which makes its willingness to hand the model unsupervised access to production systems notable. OpenAI’s own announcement frames the release around judgment as much as raw capability, citing improved design and writing quality alongside the coding and cybersecurity gains.

The Pricing Backdrop

The release lands as ChatGPT Plus, Claude Pro and Perplexity Pro all sit at the same $20 monthly price, roughly €18 to €23, meaning OpenAI’s edge now has to come from capability rather than cost. If computer-use features spread to other tiers, the practical differentiator between these subscriptions may shift from search quality to how much autonomous work a model can be trusted to do unsupervised.