What Perplexity is delegating
Perplexity says it is using GPT-6 Astra for work that extends beyond code generation: drafting communications, editing real systems and monitoring production software. In OpenAI’s customer story, Perplexity cofounder Johnny Ho says the company can trust Astra with complete end-to-end systems and check in less frequently than with earlier models.
The interesting shift is operational. Perplexity is describing the model as part of a workflow that touches live software and the surrounding systems, rather than as a coding assistant that stops at producing a patch.
Testing is the important part
One concrete workflow is automated testing. Ho says he asks Astra to build a small testing program around an application and simulate the responses that external services might return. The generated harness can then exercise the workflow from start to finish and observe how the application behaves.
That pattern reduces one of the main weaknesses of agentic coding: code can look plausible without being demonstrated to work. A model that can create the test environment, execute the workflow and inspect the result is more useful than one that only writes implementation code.
How strong is the evidence?
The claims come from an OpenAI customer story and Perplexity executives, so they should be read as first-party deployment evidence rather than an independent benchmark. The article does not publish failure rates, sample sizes or a reproducible evaluation protocol for the end-to-end workflows.
OpenAI’s broader Astra materials do provide separate benchmark evidence for stronger computer use and software-engineering performance, but that does not independently validate Perplexity’s production claims.
Why it matters
If this pattern holds up in wider use, the competitive boundary for coding agents moves from generating code toward owning verification and operation. The valuable unit becomes a completed, checked workflow rather than a diff. That is a more demanding standard, but also a much more useful one.