What AI model does it use?

What AI model does Perplexity use?

Perplexity is an answer engine rather than a lab, so most of what it runs comes from other companies. Its help center keeps a dated list, and the one I read on September 25, 2026 is labeled "a snapshot as of September 4, 2026". Perplexity says plainly that it "is not a commitment to keep any individual model available", which fits how often that list has changed.

The default: Best

Most searches run on Best. The help page is careful to say it isn't a model: "Best is not a separate mode. It is the automatic model option inside Search: Best picks a suitable model for each query." On the Free plan it's all you get: "you cannot choose an advanced Search model manually."

Which models Best chooses between is left unsaid, as is how often it lands on Perplexity's own.

The models Pro and Max can pick

Paid plans get a model selector. On September 4 it held these:

Enterprise Max gets all ten, with admins able to switch providers off. "Thinking" appears in the picker too, but Perplexity calls it "a setting on a model, not a separate model."

What runs Deep Research

Deep Research, the mode that writes a long report with many sources, doesn't use the picker. "Perplexity selects the models automatically," the models page says. A separate Deep Research article gets more specific: "Max subscribers get Opus 4.6 Thinking" and "Pro subscribers get 4.5 Thinking gradually."

Two caveats on that. The Pro line doesn't say whose 4.5 model it means, and I've left it at that rather than guess. And the article carries no date, while the September 4 search list already offers Claude Opus 5, so the Deep Research page may simply be behind.

Sonar, Perplexity's own model

Sonar is the model Perplexity makes itself, and its first version wasn't trained from scratch. Perplexity's blog post introducing it says: "Built on top of Llama 3.3 70B, Sonar has been further trained to enhance answer factuality and readability for Perplexity's default search mode." So the in-house model started as Meta's open Llama.

The picker now lists Sonar 2. I couldn't find a Perplexity announcement or spec sheet for it, so what it's built on isn't on the record. For developers, the Sonar API now sits under a "Legacy API" heading in Perplexity's docs, next to a newer API that routes to outside models.

What Perplexity doesn't say

Perplexity hasn't said which model Best uses for a given question or what Sonar 2 is built on, and hasn't dated its Deep Research page. The picker in your own account is, in Perplexity's words, "always the source of truth".