Which AI is Perplexity based on?

The Confusing Question Behind Perplexity's Answers

Ask three people what AI powers Perplexity, and you'll likely get three different answers — GPT, Claude, maybe "their own thing." The confusion is understandable, and it's the reason so many online explanations contradict each other.

Most people meet Perplexity after using ChatGPT or Claude, so they assume it works the same way: one big model, one company's brain behind every answer. But Perplexity's own documentation says it "combines web search with models built by Perplexity and by leading third-party providers." That's a fundamentally different design, and it explains why older articles list models like GPT-4o and Claude v4.0 while newer ones mention GPT-5.6 and Claude Sonnet 5 — the lineup keeps changing.

This matters for two practical reasons. First, if you're choosing a research tool, you need to know what you're actually paying for: Perplexity's Help Center lists 10 models across its paid plans, drawn from eight companies including OpenAI, Anthropic, Google, xAI, Moonshot AI, Z.ai, and NVIDIA. Second, it changes how you interpret what Perplexity tells you. As one technical comparison puts it, "Perplexity's reasoning is largely inherited from whichever LLM it employs for a query" — meaning answer quality can shift depending on which model handled your question.

The scale behind that orchestration is easy to underestimate. Perplexity reported processing 340 million queries per month as of October 2024, with more than 1,500 organizations integrating it into their workflows.

So what's the short answer?

• Perplexity is an orchestrator, not a single model — a platform that routes queries across multiple AIs with an integrated search engine, as independent analyses describe it.
• Its default backbone is Sonar, a proprietary model family fine-tuned from open-source Llama and Mistral foundations using NVIDIA's NeMo training stack.
• Everything runs through a retrieval-first architecture that grounds answers in live web search before any model generates a word.

That last point is the real differentiator. One widely shared analysis frames Perplexity as "AI that investigates" versus traditional LLMs as "AI that generates" — it dynamically picks among models per query rather than relying on one.

If that "evaluate and swap components, not magic boxes" mindset appeals to you, it applies beyond models, too. SkillzBench is building a free directory and MCP connector for doing exactly that with agent skills — discovering, verifying, and benchmarking them so you know which ones genuinely help.

Sonar Plus 9 Other Models: What Actually Powers Perplexity

So which AI actually answers when you type into Perplexity? The honest answer: it depends on your plan, your settings, and sometimes the question itself.

Perplexity's official Help Center lists 10 models across its plans, drawn from 8 companies. The lineup includes Perplexity's own Sonar 2 alongside GPT-5.6 Terra and GPT-5.6 Sol (OpenAI), Gemini 3.7 Flash (Google), Claude Sonnet 5 and Claude Opus 5 (Anthropic), Kimi K3 (Moonshot AI), GLM 5.3 (Z.ai), Grok 4.6 (xAI), and Nemotron 3 Ultra (NVIDIA).

The proprietary Sonar family is the default backbone for both free and paid tiers. According to an NVIDIA case study, Perplexity built Sonar by fine-tuning open-source models from the Llama and Mistral families using NVIDIA's NeMo training stack — scaling from 0.5B to over 400B parameters. Perplexity AI Research Engineer Weihua Hu credits NeMo with enabling rapid fine-tuning of embedding models that "greatly enhanced our retrieval stack."

Access is plan-dependent. Free users get "Best" mode with limited Pro Searches, while manual model selection requires a paid plan — and Max plans add GPT-5.6 Sol and Claude Opus 5 on top of Pro.

A few things worth knowing about how this lineup behaves:

• "Best" isn't a model — it automatically picks a suitable model for each query.
• Third-party models behave differently in Perplexity than in their native apps, because Perplexity wraps them in its own search, citations, prompts, and safety systems.
• The lineup changes frequently — Perplexity retires older models as providers ship new ones, so the model selector is the source of truth.

This multi-model approach reflects a broader shift in how AI products are built. As one analysis of Perplexity's Model Council feature put it, the industry is moving from "one assistant" toward orchestrated systems — treating models like components you evaluate, monitor, and swap rather than magic boxes.

That same mindset applies beyond models. If you're extending AI workflows with agent skills, SkillzBench is building a free directory and MCP connector that applies exactly this discipline — security-scanning, validating, and benchmarking every skill so you can equip your agents with confidence.

Why Retrieval, Not the Model, Is Perplexity's Real Edge

Ask Perplexity a question twice with different models selected, and you can get noticeably different answers — because the model is only half the story. The real engine behind Perplexity's accuracy is what happens before any model writes a single word.

Perplexity's answer pipeline begins with real-time web search, filtering, semantic ranking, and deduplication — all completed before the model is prompted. As one architectural analysis puts it, the model acts as a "controlled synthesizer bound by the available evidence." That's retrieval-augmented generation (RAG) in practice: the retrieval layer does the investigative work, and the model compresses findings into an answer.

This is why citations appear woven directly into the response through explicit context markers rather than appended after generation. The model isn't recalling facts from training data — it's synthesizing sources the retrieval layer just fetched. Perplexity itself notes that third-party models behave differently inside its platform than in their native apps, because Perplexity wraps them in its own search, citations, tools, prompts, and safety systems, per its official Help Center.

The retrieval layer isn't static, either. Perplexity fine-tunes its own embedding models to improve what sources get surfaced — a NVIDIA case study quotes AI Research Engineer Weihua Hu saying this "greatly enhanced our retrieval stack and led to a significant boost in answer quality." The same case study reports 340 million queries processed monthly and fine-tuning that scaled from 0.5B to over 400B parameters.

Why the same question yields different answers by model:

• "Best" mode routes each query to whichever of the 10 available models it judges suitable — it's not a single model at all.
• Each model synthesizes the same retrieved evidence in its own style and depth, so reasoning quality varies.
• The Model Council feature makes this explicit, running one question across three models in parallel and comparing outputs.

The takeaway from comparative analysis is clear: Perplexity's strength is information retrieval, not raw model size. The trade-off is conversational continuity — queries are handled one at a time with a fresh search, so long back-and-forth chats can feel less coherent than in ChatGPT or Claude.

That "evaluate, monitor, and swap components" mindset extends beyond models. If you build AI workflows, agent skills deserve the same scrutiny — SkillzBench is building a free directory and MCP connector that benchmarks skills in a sandbox so you can see which ones actually help. Join the waitlist to go further.

How to Choose and Use Perplexity's Models Like a Pro

Knowing which models sit behind Perplexity's answers only helps if you actually use that knowledge — and the good news is Perplexity gives you the levers to do it deliberately rather than accepting whatever "Best" mode hands you.

Start with the model selector, and treat it as your source of truth. Perplexity's own Help Center notes that it adds new models and retires older ones as providers release them, so any article (this one included) can be stale by the time you read it. A lineup that once featured GPT-4 and Claude v4.0 now lists ten models from eight companies, including Kimi K3 and Nemotron 3 Ultra, according to Perplexity's official documentation.

Next, use Model Council when a question matters enough to cross-check. The feature runs one query across three models in parallel and synthesizes the results in a comparison table, as reporting on the launch explains. It was Max-only and web-only at launch, so check your plan — but for high-stakes research, seeing where models agree and diverge is worth the extra step.

Finally, match models to tasks. Remember that analysis of Perplexity's architecture found its reasoning is largely inherited from whichever LLM it employs for a query, so the model you pick shapes the answer you get. A quick mental checklist:

• Fast, cited summaries: start with Sonar, which is purpose-built for web-grounded responses.
• Complex reasoning or analysis: pick a frontier model with stronger inference.
• Contested or high-stakes questions: run Model Council and compare three answers.
• Everyday queries: let "Best" mode pick — it routes automatically per query.

The broader lesson extends well beyond Perplexity. As gend.co put it, treat models like components you evaluate, monitor, and swap — not magic boxes. That same mindset applies to the agent skills layer of your AI stack, where published research found roughly 26% of skills carry at least one vulnerability. SkillzBench is being built precisely for this: a free directory and MCP connector that pin, scan, and benchmark skills so you can adopt them with evidence rather than hope. If evaluating AI components sounds like work, that's because it is — but it's the difference between guessing and knowing.

Evaluate, Don't Trust: The SkillzBench Way to Vet AI Components

Perplexity's own product decisions reveal something important: its Model Council feature runs a single question across three models in parallel and synthesizes the results, an acknowledgment from Perplexity itself that different models excel at different tasks. The lesson generalizes far beyond chatbots. If even a leading AI company refuses to trust one component blindly, anyone building AI agents should apply the same discipline to every part of their stack — including the agent skills their tools depend on.

Agent skills — SKILL.md instruction packages used by Claude Code, Codex, Cursor, Gemini CLI and chat-app agents — have spread quickly across GitHub and skill marketplaces. But speed has outpaced scrutiny. Published research found that roughly 26% of agent skills carry at least one vulnerability, meaning about one in four components you install could expose your agent, your data, or your workflow to risk. Treat skills like magic boxes and you inherit problems you never saw coming.

The better approach is the one Perplexity embodies: evaluate, monitor, and swap — never assume. That means asking concrete questions before a skill touches your agent:

• Is the code secure? Has it been scanned for known vulnerabilities?
• Does it actually improve results, or just add complexity?
• Who maintains it, and does the license allow your use case?
• Is the version you tested the version you're running?

Most teams can't answer these questions manually — there are too many skills and too little time. That's the gap SkillzBench, AIQ Labs' free and independent directory of agent skills, is being built to close. Its approach mirrors the orchestration mindset: every skill is pinned to a commit SHA, security-scanned, license-checked, and benchmarked in a sandbox with the same model, run with and without the skill, so you can see the Lift in percentage points rather than trusting a description. A trust score from 0 to 100 combines security, quality, license, maintenance, and community signals — and Lift is always shown separately, because money never touches scores.

The platform is still in active development, but the core pieces are taking shape: a continuously crawled and deduplicated skills directory, sandbox benchmarking, ongoing re-scanning of upstream changes, and a free MCP connector that will let any agent load only the skills you explicitly choose. Everything — directory, benchmarks, alerts, and the MCP tool — will be free, with revenue coming from ads around blog content rather than from users.

Perplexity doesn't ask you to trust a single model; it grounds answers in evidence and lets components prove themselves. Your agents deserve the same standard. If you want skills that are proven rather than promised, join the SkillzBench waitlist and be first in line when the directory and free MCP connector go live.