Perplexity Hands GPT-6 Astra the Keys to Its Production Systems

Perplexity, the AI search company, says it has moved past using GPT-6 Astra for coding assistance and now trusts the model to “edit real-world systems, and monitor our production software” with reduced human checking. Cofounder and chief strategy officer Johnny Ho put it directly in OpenAI’s September 14 case study: “we’re actually able to trust it with full end-to-end systems and check in on it much less frequently than previous generations of models.” One concrete workflow: Perplexity uses Astra to build small testing programs that generate realistic mock responses — standing in for a language-model API or a connector — so an application’s behavior can be validated end-to-end without a person running the tests by hand.

The trust claim lands eleven days after OpenAI’s own GPT-6 Astra launch, in which the company disclosed the model crossing a “Critical” cybersecurity capability threshold under its Preparedness Framework — strong enough to independently discover and exploit unknown vulnerabilities in well-protected systems, and gated behind OpenAI’s Daybreak program as a result. That combination is the real story for anyone evaluating frontier-model adoption: the same capability jump that lets Perplexity delegate production monitoring is explicitly the one OpenAI felt needed slow, gated rollout for offensive security use.

It also isn’t the only production-grade trust signal Astra has picked up this month. Cognition, the company behind the autonomous coding agent Devin, published its own case study describing how Astra now tests and proves Devin’s own work — returning screen recordings and pass/fail reports rather than just a diff — with cofounder Walden Yan framing the goal as engineers who “manually look at less code and end up shipping more.” Two frontier-adjacent AI companies, in the same two-week window, both describing reduced human oversight as the payoff of the same model generation — a pattern enterprise buyers evaluating agentic coding tools should treat as more than marketing.