What does Perplexity use for AI?

Why 'Which Model Does Perplexity Use?' Is the Wrong Question

Ask "which model does Perplexity use?" and you're asking the wrong question — because there's no single answer. Perplexity isn't one AI model; it's an answer engine that orchestrates more than 20 different models simultaneously, all layered over a live web search index. The model is just the last step in a much longer pipeline.

That distinction matters more than most businesses realize. Perplexity's real competitive advantage, according to an engineering analysis of its architecture, is "not any single LLM, but the proprietary orchestration system that manages the interaction with these models." When you type a query, small classifier models first parse your intent, then route the request to the smallest model that will still deliver the best experience — a simple definition goes to a lightweight in-house model, while complex multi-step reasoning gets escalated to frontier models from OpenAI or Anthropic.

Under the hood, the system runs on a retrieval-augmented generation (RAG) pipeline. Perplexity maintains its own search index tracking over 200 billion unique URLs, backed by tens of thousands of CPUs and more than 400 petabytes of hot storage, with tens of thousands of index updates processed every second. Its guiding principle for answers is strict: "you are not supposed to say anything that you didn't retrieve." The infrastructure serving all of this runs on NVIDIA H100 GPUs and the Triton Inference Server, handling over 400 million queries per month, according to Perplexity's own engineers.

For business readers evaluating AI tools, this architecture reveals three things worth checking before you trust or adopt any AI platform:

• Is the system grounded in live, retrievable data, or does it lean on static training data that goes stale?
• Does it route intelligently between models, or force every query through one heavyweight LLM regardless of cost?
• Can it show you where each claim came from — and how reliable those sources actually are?

That last point deserves a caveat. Independent testing found a 37% citation error rate in Perplexity's answers — the best result among eight AI search tools tested, where over 60% of all responses across tools contained citation errors, per comparative testing. Best-in-class still isn't perfect, which is why treating any AI answer engine as an intelligent gateway rather than a final authority remains the smart default.

The same verification-first mindset applies beyond search. Agent skills — instruction packages that extend coding agents and chat-app agents — are multiplying fast, but published research found roughly 26% carry at least one vulnerability. That's the problem SkillzBench is building against: a free directory that scans, benchmarks, and trust-scores agent skills so you can see which ones actually help before you wire them in. Understanding what powers a tool — and verifying it independently — is quickly becoming a core business skill, not a technical nicety.

The Models Behind Perplexity: Sonar, GPT, and Claude with Smart Routing

Ask Perplexity a question and you're not talking to one AI model — you're talking to a small committee of them, coordinated by a routing system that decides which brain handles your query. That orchestration layer, not any single model, is what engineering analyses consistently identify as the company's real "moat."

Perplexity runs a deliberately mixed model fleet. On one side sits its in-house Sonar family — fine-tuned models built atop open-source foundations like Llama, optimized for grounded, citation-heavy answers. On the other side are frontier models from OpenAI (GPT) and Anthropic (Claude), reserved primarily for Pro and Max tier users who need heavier reasoning, according to technical breakdowns of its stack.

The scale is striking. Perplexity's own engineers report serving more than 20 models simultaneously — including Llama 3.1 variants at 8B, 70B, and 405B parameters — while handling over 400 million search queries per month, per a NVIDIA Developer Blog post co-authored by the Perplexity inference team.

The key economic trick is query routing. Small classifier models first parse intent, then send each request to the most cost-effective model capable of handling it:

• Simple definitional queries route to small, cheap in-house models
• Complex multi-step reasoning escalates to frontier models from OpenAI and Anthropic
• The guiding principle: "the smallest model that will still give the best possible user experience"

This routing discipline pays off in hard numbers. Tensor parallelism across 4–8 NVIDIA H100 GPUs delivers up to 3x lower relative cost per million tokens for latency-sensitive requests, and Perplexity saved roughly $1 million annually by moving its Related-Questions feature from third-party LLM APIs to self-hosted NVIDIA GPUs running Triton Inference Server and TensorRT-LLM.

The lesson generalizes beyond search. Model capability is commoditizing fast; the differentiation is shifting to orchestration — choosing the right tool per task, grounding outputs in real data, and verifying results. Independent testing found Perplexity still produces a 37% citation error rate, the best of eight AI search tools tested but far from flawless, per independent comparisons. Even the best orchestration layer needs verification downstream.

That same verify-before-you-trust logic applies to the agent-skills ecosystem: published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free directory that pins, scans, and benchmarks agent skills so builders can measure whether one actually helps — the same "prove it, don't assume it" philosophy Perplexity applies to model selection.

The Data: A 200-Billion-URL Live Index and the RAG Pipeline

Beneath every Perplexity answer sits a retrieval machine most users never see: a proprietary index of more than 200 billion unique URLs, refreshed with tens of thousands of updates every second. That live index — not the language model — is what makes Perplexity feel current.

Perplexity started in 2022 leaning on the Bing Web Search API, but it has since built its own search infrastructure entirely, with no static cached answer store. According to engineering analysis, the index is supported by tens of thousands of CPUs and over 400 petabytes of hot storage, maintained by a team of roughly 38 engineers.

The retrieval layer runs on Vespa AI, an engine chosen for real-time RAG at scale because it combines vector search, BM25 lexical search, structured filtering, and machine-learned ranking in one system. A pipeline breakdown describes a six-stage process:

• Query intent parsing to classify what the user actually wants
• Embedding-based indexing of live web content
• Multi-method retrieval — standard search pulls 60+ sources per query
• Multi-layer ML ranking of retrieved results
• Structured prompt assembly with citations pre-embedded
• Constrained LLM synthesis into one cited answer

Since February 2025, Perplexity has also used its own pplx-embed embedding models (0.6B and 4B parameters, built on Qwen3), replacing third-party providers. The 4B context model scores 81.96% on the ConTEB benchmark, beating Voyage's 79.45%.

The grounding principle is strict: as Perplexity puts it, you are not supposed to say anything you didn't retrieve. Perplexity would reportedly rather return nothing than serve weak citations, restarting retrieval if too few results pass quality thresholds.

Data sources go beyond raw crawling. Perplexity runs a publisher revenue-sharing program with outlets like TIME, Fortune, Der Spiegel, and the LA Times, while some publishers — including The New York Times — have blocked its crawlers or entered litigation. Enterprise Pro adds private corporate corpora.

Even this best-in-class retrieval isn't flawless. The Columbia Journalism Review's testing found a 37% citation error rate in Perplexity's answers — the best of eight AI search tools tested, but a reminder that retrieved sources still need human verification. That same verification mindset drives SkillzBench, which is building a free directory and MCP connector that security-scans, validates, and benchmarks agent skills before you rely on them.

What This Means for Your Business: Trust, Verify, and Build on Top

If Perplexity itself can't always be trusted to cite correctly, your business shouldn't treat it as gospel. Independent testing by the Columbia Journalism Review found a 37% citation error rate in Perplexity's answers — still the best of eight AI search tools tested, but far from flawless. That's the right lens for using it at work: an intelligent gateway, not a final authority.

DataStudios puts the best practice plainly for news and current events: treat Perplexity as an intelligent gateway, not a final authority. In practice that means reading the cited sources behind any answer before you act on it, especially for research that feeds decisions, briefs, or published content. The good news is that Perplexity makes verification easy — every answer includes clickable citations, and standard search retrieves 60+ sources per query, with Deep Research reading hundreds.

So where does Perplexity genuinely fit into business workflows?

• Rapid market and competitor research, where speed and breadth of sources matter more than perfect citation accuracy.
• News and trend monitoring, using its live index of over 200 billion URLs rather than static training data.
• Automated research pipelines via its API, with Sonar models starting at $1.00 per million tokens for input and output.
• Enterprise knowledge search, combining the live web with private company corpora.

The self-reported numbers from its API platform are encouraging if you verify them against your own use case: Copy.ai reports companies saved 8 hours of research per rep per week, translating to a 20% throughput increase, while Apollo users report 46% more booked meetings. Treat these as vendor claims worth piloting, not guarantees.

The deeper lesson here applies well beyond Perplexity. Its own engineering philosophy — "you are not supposed to say anything that you didn't retrieve" — is a verification discipline worth copying. And if your team is building agents that combine search-augmented AI with skills and tools, the same trust-but-verify principle matters even more, because agent skills are largely unvetted. Published research found roughly 26% of agent skills carry at least one vulnerability, yet there's been no easy way to tell which ones actually help.

That's the gap SkillzBench is building to close: a free, independent directory of agent skills with a free MCP connector, where every skill is version-pinned, security-scanned, license-checked, and benchmarked in a sandbox — with the performance Lift published separately from any ratings. It's currently in pre-launch, with the waitlist live. Whether you're verifying an answer engine's citations or equipping an agent with new skills, the principle is the same: trust, but verify — and build on top only once you know what's underneath.