Choosing the Right Perplexity Model

Part of a series on running AI workflows across model providers — this one is about picking a tier without defaulting to the most expensive "to be safe."

Four tiers, two naming schemes in transition, and a genuine cost difference between the cheapest and most expensive option. Picking the right one isn't complicated once you know what each tier is actually built for — the mistake most teams make is defaulting to the most capable tier "to be safe," which mostly means paying for search depth a simple question never needed.

The four tiers, old names and new

Perplexity's own documentation describes the lineup this way, and the same four tiers carry over into the new Agent API as presets (covered in the previous two articles):

There's also xhigh, new to the Agent API with no direct Sonar predecessor, for open-ended agentic research beyond what Deep Research was scoped for.

Complexity and cost both climb left to right — the question is how far right a given task actually needs to go.

A three-question framework

Does this need current information at all, or just reasoning? If the question doesn't depend on anything that happened recently, you may not need Perplexity at all — a purely parametric model from another provider in this series might answer faster and cheaper. This is worth checking first, before picking a Perplexity tier.

Is one search pass enough, or does this need follow-up digging? A single factual lookup ("what's the current status of X") fits the lightest tier fine. A question that requires cross-referencing several sources, resolving contradictions between them, or following one finding to the next research question needs one of the reasoning or research tiers.

Is this a single answer, or a report? If the deliverable is a short, direct answer, the top research tier is overkill — it's built to produce comprehensive, multi-source reports, not quick responses, and it's priced and structured for that job. If the deliverable genuinely is a report — competitive analysis, due diligence, a literature-style synthesis — that's exactly what it's for.

Testing this without guessing

DNotifier's Prompt Testing Studio applies here the same way it applies to every other provider in this series: run the same prompt against multiple tiers side by side and compare the actual answer quality, latency, and cost before committing production traffic to any one of them. This turns "which tier does this need" from an educated guess into a five-minute comparison with real numbers attached.

Testing three tiers against the same question replaces a guess with a side-by-side comparison.

A use case that shows the cost of over-provisioning

A customer support tool was routing every question that mentioned a product name through the top research tier, on the assumption that "more thorough" was always the safer default. A Prompt Testing Studio comparison showed the lightweight tier produced functionally identical answers for roughly 80% of those questions — simple "is this feature available" and "what's the current price" lookups that didn't need multi-source synthesis at all. Reserving the research tier for the genuinely complex subset (competitive comparison questions, "how does this compare to three alternatives" type requests) cut the average cost per support query considerably without any drop in answer quality on the simple majority.

The one thing this framework doesn't cover

None of the four tiers should be the default choice if the question doesn't need current information at all. It's easy to reach for Perplexity out of habit once it's connected and working well, but a purely parametric model — from any other provider in this series — is often faster and cheaper for reasoning-heavy or coding-heavy tasks that don't hinge on anything time-sensitive. Tier selection matters, but provider selection comes first.

Frequently asked questions

Is the most expensive tier always the most accurate?

Not necessarily for simple questions — it's built for depth and breadth of research, which isn't the same axis as correctness on a straightforward factual lookup. Test before assuming more expensive means better for your specific use case.

What happens to my tier choice after the September 27 Sonar retirement?

The mapping in the previous article carries directly: Sonar → fast, Sonar Pro → low, Sonar Reasoning Pro → medium, Sonar Deep Research → high. The decision framework in this article applies unchanged to whichever naming scheme is current.

Can I switch tiers per-request rather than committing to one for an entire application?

Yes — since tier selection happens through the model (or preset) parameter on each sendAI() call, different steps in the same DNotifier workflow can use different tiers based on what each step actually needs.

Does a heavier tier always mean slower responses?

Generally yes, since deeper research means more search passes and more synthesis before an answer is ready. That's part of why matching the tier to the actual task matters for user-facing latency, not just cost.

The Bottom Line

Next in this series: what these tiers actually cost, and how the pricing model changes once the Agent API takes over.

Read part six: Perplexity Pricing, Explained. Explore dnotifier.com.