Perplexity trusts GPT-6 Astra with end-to-end production systems

The case study is part of OpenAI's push to demonstrate that Astra is production-ready for high-stakes agentic work, not just chat. Perplexity says it uses the model to write internal communications, modify software, and monitor production systems, and — the key claim — checks in far less frequently than with earlier models, indicating the agent runs longer autonomous loops with less human supervision.

That 'trust' framing is the strategic thrust: OpenAI is arguing Astra crosses a reliability threshold where engineers can delegate monitoring of live systems. Paired with a companion case study on Cognition's Devin using Astra to test its own work, OpenAI is building a narrative that Astra closes the self-verification loop that has held agents back from production.

The context makes this double-edged. The same week, OpenAI's own Astra downgrade complaints surfaced on Reddit, and Anthropic's threat report documented agents being weaponized — so 'less human oversight of production systems' cuts against the industry's simultaneous safety anxiety. Whether trusting an agent to monitor production is prudent efficiency or premature delegation is exactly the tension Amodei's slowdown essay is about.