AI Agent · Perplexity
Behind the search bar,
the agent's brain just changed.
Perplexity has switched the backbone model of its autonomous execution mode, "Perplexity Computer," to OpenAI's newest model, GPT-6 Astra. Rather than sitting alongside GPT-5.6 and Claude Opus 5 as just another selectable option, it now ships as the default model for Pro/Max users — here's what's actually changing under the hood of a search AI.
The Switch
Not about speed —
about what it can be trusted with
The switch wasn't driven by raw speed. It was driven by how much unsupervised work the model can now be trusted with.
On September 3, 2026, OpenAI released GPT-6 Astra to approved users, opening it to general availability the following day. Perplexity co-founder Aravind Srinivas reacted almost immediately, posting on X that the model was "far ahead of every other model on wide and deep research tasks, while also being more cost-effective," and announcing plans to bring it to Perplexity Computer for Pro and Max users.
True to that word, the official @perplexity_ai account confirmed around September 10 that "GPT-6 Astra is now available in Perplexity Computer for Pro and Max subscribers." Srinivas followed up, noting that "Pro and Max subscribers get to use both Fable and Astra on Computer mode." In other words, the existing model choices stayed intact — only the default backbone was swapped for Astra. The same arc is corroborated by OpenAI's own case study, "Perplexity trusts GPT-6 Astra with end-to-end systems".
Worth flagging: "Perplexity Computer" refers specifically to the autonomous agent mode that operates a browser or files on its own — it is a separate surface from the ordinary one-shot answer box. The core search experience itself hasn't switched to Astra outright, a distinction that's easy to miss.
The numbers behind the pick
On the same internal benchmark, GPT-6 Astra scored 27.0% higher than Claude Opus 5, for only 3.3% more in cost. That said, this is Perplexity's own benchmark design and reporting — not an independently reproduced, third-party evaluation — so the numbers should be read with that caveat in mind.
"We can have the model craft communications, edit real-world systems, and monitor our production software in a way that previous generations were not able to."
— Johnny Ho, Cofounder and Chief Strategy Officer, Perplexity
Who It Changes
Who feels this, and how
Engineers
More autonomous-agent accuracy means more code changes and production monitoring can be delegated. But the less often a human checks in, the more a weak sandbox design turns straight into an incident.
Business / PM
Anyone weighing whether to delegate internal system maintenance or reporting to an agent now has concrete cost-vs-accuracy numbers ($11.98/task, etc.) to reason from, rather than a vague sense of "it's better."
Everyday users
For ordinary search and Q&A, the difference is barely noticeable. Unless you're actually using Computer mode, day-to-day usage is largely unaffected.
What to watch next
Test cost on real workloads
Pro/Max users should toggle between Astra and Fable 5.1 in Computer mode's settings and measure speed, accuracy, and cost against their own workload before committing to one as the default.
Watch for copycat moves elsewhere
Making a frontier model the default engine behind your own autonomous product is a pattern likely to spread to other search and productivity tools. Watch for similar switch announcements over the coming weeks.
Codify the rules before checking in less
"You can check in less often" cuts both ways — it also means failures a human would have caught can slip through unnoticed. Decide how much autonomy to grant before, not after, putting this into production.
Caveats
The optimistic read isn't the whole story
Because the numbers come from Perplexity's own internal evaluation, the claim that "Astra is the best" should be taken with a grain of salt — no independent third party has reproduced it yet. It's also worth noting that, just days earlier, reports surfaced of a hacking campaign abusing OpenAI's separate Codex product to breach more than ten additional sites. The very design choice of handing agents broad permissions while reducing human oversight carries a security cost that's easy to lose sight of amid the performance numbers. Before chasing speed and accuracy gains, it's worth thinking through how to bound the damage when something goes wrong.