Perplexity Says It Trusts GPT-6 Astra With Production Systems

Perplexity says it is trusting GPT-6 Astra with work that reaches beyond coding suggestions: writing communications, changing software and monitoring production systems. The consequential shift is not simply the range of tasks. Perplexity cofounder Johnny Ho says the company can check on the model’s work much less often than it did with earlier generations.

A model moves closer to the running system

That is a meaningful boundary for an AI deployment. A model can be useful when it drafts an isolated program; it takes on a different role when its output is used to alter real software and watch the systems already serving customers. Perplexity’s description places Astra in both settings, though it does not detail which production actions the model may take without a person’s approval.

Where Perplexity says it uses Astra

• Crafting communications for the company.
• Changing software and monitoring production systems.
• Building programs that test application workflows from beginning to end.

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.

Johnny Ho, Cofounder and Chief Strategy Officer, Perplexity

Testing by impersonating outside services

Ho identifies software testing as one of the most useful applications. Rather than manually test every case, he says Perplexity asks Astra to build a small testing program around an application. The model can generate plausible replies from services the application depends on, such as a language-model API or a connector.

Those simulated services let the test observe how an application responds across a complete workflow. The approach is more than asking a model to inspect a code file: Astra is being used to create the surrounding conditions that can expose whether connected parts of a program work together.

Better code is also a search capability

For Perplexity, the link between model coding ability and its core product is direct. Ho says stronger code generation helps the company’s search engine write better programs to search the web and internal information, then summarize the results concisely. Astra’s value, in this account, is therefore not limited to developer productivity; it can improve the software that performs the search work itself.

Trust is the unresolved part

The disclosure is a customer account published by OpenAI and reflects Perplexity’s own assessment of the model. It does not provide error rates, time-saved figures, production-incident data or a precise description of when human review remains mandatory. That leaves the central operational question open: whether less frequent checking is a reliable result of better testing, or a judgment each company must continually revisit as an AI system gains access to more consequential work.