Perplexity hands Astra end-to-end system work
OpenAI's own customer page says the model drafts communications, ships code changes and watches production, with far fewer check-ins.
In short
Perplexity uses GPT-6 Astra for whole workstreams — drafting communications, changing software and monitoring production systems — and, per OpenAI, reviews the model's work far less often than it did with earlier models.
At a glance
- Three named uses: drafting communications, changing software, monitoring live production systems.
- Perplexity reviews the model's work far less often than with earlier models, according to OpenAI.
- The only source is an OpenAI customer page: vendor communication, not independent reporting.
- That page could not be read while writing this article (HTTP 403); no figures or measurements are available.
Perplexity has moved GPT-6 Astra from suggestion duty to owning work end to end, according to a customer story published by OpenAI. The model drafts communications, makes changes to software, and watches systems that are live. The notable part is not the task list but the supervision rate: the team reportedly checks in far less often than it did with earlier models.
What the report says
Three areas are named: writing communications, changing software, and monitoring production systems. Each already existed as an assisted task. What changes is who holds the intermediate steps.
When a model only proposes, a person approves every hop. When a model owns the chain, review moves to the end of it, and the number of review points drops by design.
Why the check-in rate is the story
Fewer check-ins means a mistake survives longer before anyone notices. In production that is a question of downtime and data integrity, not convenience. A company taking this step either has strong rollback paths or has accepted more exposure. The report does not say which.
What is not established
The only available source is OpenAI's own customer page, which is vendor communication about a customer rather than independent reporting. At the time of writing that page returned HTTP 403 and could not be read, so this article relies on the short excerpt captured when the item was collected.
That leaves several things unverified: how long the setup has been running, how many systems it covers, any measured error or incident rate, and how often humans still step in. We will add figures once a second, independent newsroom covers it.
What to take from it
For engineering teams the transferable question is operational, not which model to pick. Agents that touch production need audit logs, a rollback path, and a named owner for the moment something breaks. Those requirements do not disappear because the review step got rarer.
FAQ
What does Perplexity use GPT-6 Astra for?
Per OpenAI, three areas: writing communications, changing software, and monitoring production systems, each handled as a full task rather than a suggestion.
Does Astra run production systems without supervision?
The source does not say that. It only states that check-ins are far less frequent than with earlier models; how often people intervene is not given.
How reliable is this report?
Limited. It is an OpenAI customer page, so it is vendor communication. No second, independent newsroom has confirmed the account so far.
