Perplexity hands GPT-6 Astra day-to-day control of its production systems

Perplexity has started using OpenAI’s GPT-6 Astra model to manage core parts of its production systems, according to OpenAI. The company says Astra writes communications, changes software, and monitors the infrastructure that keeps Perplexity’s products running, and that staff check in on its work far less often than they did with earlier models.

OpenAI’s description groups Astra’s work at Perplexity into three tasks. It drafts communications, a category OpenAI does not further specify, meaning it is unclear whether that covers internal updates, customer-facing messages, or both. It changes software, which in production engineering means pushing code updates to systems already serving users. And it monitors production systems, industry language for the live infrastructure that keeps an application running and available.

The detail OpenAI emphasizes most is the drop in oversight. Companies that let AI models touch live infrastructure typically keep a human reviewing each action before or shortly after it happens, precisely because a bad deployment or a missed alert can take a product down. OpenAI’s account frames Astra as needing that kind of review less often than the models Perplexity used before it, though OpenAI does not say what the earlier check-in frequency was, what it is now, or what specific errors prompted the tighter process in the first place.

What’s attributed and what isn’t

Every claim here traces back to OpenAI describing its own model’s deployment at a customer, not to an independent account from Perplexity or a third party. OpenAI is the party with a direct interest in Astra’s reputation for reliability, so its description of reduced oversight should be read as OpenAI’s characterization rather than a verified operational fact.

Several things the source material does not say are worth stating directly. There is no figure for how often Perplexity now reviews Astra’s actions, no timeline for when this arrangement began, no mention of any incident tied to Astra’s changes, and no comparison to how other companies handle similar oversight questions with production-facing AI systems. Whether Astra’s software changes go through automated testing before reaching users, or whether a human still signs off on higher-risk deployments, is not addressed.

What to watch next is whether OpenAI or Perplexity publish more concrete detail on how this oversight arrangement works in practice, including any guardrails specific to code deployment and infrastructure monitoring. A public account of an incident, or the absence of one over time, would say more about how the arrangement is holding up than the current description does.