What AI Models Does Perplexity Use?

Why 'Which Model Powers Perplexity?' Is the Wrong Question

Most people asking "which model powers Perplexity?" expect a one-word answer — GPT, Claude, or Gemini. The real answer is more interesting: Perplexity doesn't run on one model at all. It's a multi-model aggregator that pairs its own Sonar family with models from seven third-party providers: OpenAI, Google, Anthropic, xAI, Moonshot AI, Z.ai, and NVIDIA, according to the company's Help Center documentation.

Perplexity describes this as routing queries "across multiple frontier models so each question is handled by the model best suited for it." When you leave the selector on the default "Best" setting, you're not choosing a model — you're letting Perplexity pick one per query. That's why the question "which model does Perplexity use?" has no single answer: it depends on what you asked, when you asked it, and which subscription tier you're on.

The second complication is that the lineup never sits still. Perplexity explicitly states its models are "subject to change" and that the in-app model selector is "the source of truth" for your account. Any list of specific models — including the one later in this article — is a point-in-time snapshot, not a permanent catalog. A few patterns from the data illustrate how fast things move:

• Perplexity made all 46 new models launched by leading AI companies in 2025 available within 24 hours of release, per its enterprise blog.
• In January 2025, two models (Claude Sonnet 4 and GPT-4o) handled 91.5% of enterprise queries; by late 2025, no single model exceeded roughly 23% share.
• 40% of power users actively work with six or more models, versus 20% of regular users.
• The top 50 enterprise accounts use 30 models on average, compared with seven for typical accounts.

There's a third wrinkle: even when you select a familiar model like Claude or GPT inside Perplexity, you're not getting the stock version. Perplexity integrates third-party models with its own search, citations, prompts, and safety systems, so the same model can behave differently there than in its standalone app. And its proprietary Sonar line has its own lineage — the original was built on Meta's Llama 3.3 70B and fine-tuned for factuality and readability, per Perplexity's Sonar announcement.

The better question, then, isn't "which model?" but "how does Perplexity match models to tasks?" That task-matching mindset — picking the right tool per job rather than defaulting to one — is exactly the logic behind agent skills, and it's what SkillzBench is building its free directory and MCP connector around: helping you find, verify, and benchmark which skills actually help your agent. Access to every option matters more than choosing today's best one — because, as Perplexity's own data shows, what leads today may not lead next quarter.

Inside the Model Lineup: Sonar, Frontier Partners, and Plan Tiers

Perplexity answers your question with ten different brains, not one. Understanding which model you're getting — and which tier unlocks it — is the difference between a $20 subscription and an enterprise contract.

Perplexity's proprietary Sonar model was built on Meta's Llama 3.3 70B and fine-tuned specifically for answer factuality and readability, according to the company's own engineering blog. The fine-tuning pipeline matters here: Perplexity used NVIDIA NeMo to apply supervised fine-tuning, direct preference optimization, and PPO alignment — the same techniques that produced a 20% quality improvement over base models within days of open-source releases.

Speed is Sonar's signature. It runs at roughly 1,200 tokens per second on Cerebras inference infrastructure — about 10x faster decoding than Gemini 2.0 Flash, per Perplexity's internal testing. That velocity exists because Perplexity was handling 340 million queries per month as of October 2024, and latency compounds at that scale.

Perplexity layers its model catalog across subscription tiers, pulling from OpenAI, Google, Anthropic, xAI, Moonshot AI, Z.ai, and NVIDIA. As of the September 2026 Help Center snapshot:

• Pro and Max: Sonar 2, GPT-5.6 Terra, Gemini 3.7 Flash, Claude Sonnet 5, Kimi K3, GLM 5.3, Grok 4.6, and Nemotron 3 Ultra — eight models total.
• Max only: GPT-5.6 Sol and Claude Opus 5, the heavyweight reasoning models reserved for the top consumer tier.
• Enterprise Pro and Enterprise Max: all ten models on both tiers, which explains why the top 50 enterprise accounts use 30 models on average versus seven for typical accounts.

One important clarification: the default "Best" setting is automatic routing, not a separate model. Perplexity's Help Center confirms it selects a suitable model per query, so users never need to pick manually.

Perplexity states plainly that models are "subject to change" and that the in-app model selector is "the source of truth for your account." The company offered all 46 new models launched by leading AI labs in 2025 within 24 hours of release, per its enterprise switching analysis. Treat any list — including this one — as a point-in-time snapshot.

That task-matching philosophy extends beyond models. The same logic applies to agent skills: pairing the right instruction package with the right job. SkillzBench, currently in pre-launch development, is building a free directory and MCP connector to let users discover, verify, and benchmark agent skills the way Perplexity users now route between models — as a menu of options, not a single default tool.

The Same Model Behaves Differently Inside Perplexity

Picking Claude or GPT inside Perplexity is not the same experience as using Claude or GPT directly. Perplexity openly acknowledges this: it integrates third-party models with its own search, citations, tools, prompts, safety systems, and usage limits, so a model in Perplexity can behave differently from the same model in the provider's standalone product.

That wrapper is the entire point. Perplexity's architecture routes queries across multiple frontier models so each question reaches the model best suited for it, and the default "Best" setting picks a model per query automatically. The underlying model is one ingredient; Perplexity's retrieval stack, prompt scaffolding, and citation layer shape the final answer just as much.

For Sonar specifically, the wrapping goes deeper. Perplexity built its family of online LLMs on open-source foundations using NVIDIA NeMo, applying supervised fine-tuning and direct preference optimization, and fine-tuned open-source embedding models to boost retrieval quality. New Sonar variants reportedly achieved a 20% improvement over their base models on search tasks within days of release. Those gains come from Perplexity's post-training, not the base model alone.

Here is the caveat most coverage skips: Sonar's headline performance claims come from Perplexity's own internal A/B testing, not independent benchmarks. When Perplexity says Sonar outperforms Claude 3.5 Sonnet and GPT-4o mini "by a substantial margin" and closely approaches GPT-4o, that is vendor-reported testing on Perplexity's own evaluation setup. No independent third party has verified it.

This matters for buyers evaluating any AI tool, not just Perplexity:

• Ask who ran the benchmark — vendor A/B tests measure performance on the vendor's chosen tasks, which may not match yours.
• Test the same model in both environments before assuming parity — the wrapper changes the output.
• Treat "outperforms X" claims as directional evidence, not settled fact, until independently replicated.

The same critical lens applies to the agent-skills ecosystem SkillzBench is built around: published research found roughly 26% of agent skills carry at least one vulnerability, and most skill claims are likewise self-reported. SkillzBench is building a free directory and MCP connector that pins, scans, and benchmarks skills in a sandbox — testing each one with and without the skill to publish a measurable Lift — so performance claims rest on evidence rather than marketing.

The takeaway: a model's name on the label tells you less than you think. What surrounds the model — retrieval, prompting, safety layers, and evaluation rigor — determines what you actually get.

What Enterprise Model-Switching Data Tells Us About Choosing Models

In January 2025, just two models — Claude Sonnet 4 (47.5%) and GPT-4o (44%) — handled 91.5% of all enterprise queries on Perplexity. By late 2025, no single model exceeded roughly 23% share, according to Perplexity's enterprise model-switching data. That collapse of concentration tells you something important: sophisticated organizations stopped picking a favorite model and started picking the right model for each task.

The usage patterns among the most engaged teams are striking. Forty percent of power users actively work across six or more models, versus 20% of regular users. The top 50 enterprise accounts average 30 models in active use, compared with seven for typical accounts. And 53% of users who deliberately choose specific models switch between them within a single workday at least once.

What does task-matching look like in practice? The same data shows clear specialization emerging:

• Claude handles programming, drawing 38% of coding-related queries
• Gemini Flash dominates visual arts work at 40% of those queries
• Gemini 3.0 Pro Thinking leads financial analysis with 31% share
• No single model holds more than about a quarter of any category

The takeaway for business leaders is straightforward: treat models as a menu of task-specific options, not a single default tool. A team that routes coding questions to Claude, visual work to Gemini, and analytical research to a reasoning-optimized model will outperform one that reflexively opens the same chatbot for everything. Perplexity's own framing captures the strategic logic — when the landscape shifts this quickly, access to every option matters more than choosing today's best one.

There's a second lesson here about speed. Perplexity made all 46 new models launched by leading AI companies in 2025 available within 24 hours of release, and the company explicitly tells users the in-app model selector is the source of truth for their account, since the lineup changes constantly. Whatever leads today may not lead next quarter, so lock in workflows — not model names.

The same task-matching principle extends beyond models to the tools and instruction packages your AI agents use. SkillzBench is building a free directory and MCP connector that applies this logic to agent skills: discover, verify, and benchmark skills so each agent gets the right capability for each job — the same discipline enterprise teams are already applying to model selection.

Match the tool to the task, and review the lineup regularly. That's the real insight hiding in the model-switching data — and it applies whether you're choosing between frontier models or equipping an agent for its next workflow.

Putting Task-Based Model Selection to Work (and Going Further with Agent Skills)

The fastest way to know exactly which models your Perplexity plan unlocks is not a blog post or a changelog — it's the app itself. Perplexity explicitly tells users that "the model selector in Perplexity is always the source of truth for your account," because the lineup is subject to change at any moment. In 2025 alone, the platform added all 46 new models launched by leading AI companies within 24 hours of release, according to Perplexity's enterprise model-switching report. Any specific list you read today is a point-in-time snapshot.

That said, the snapshots are useful for planning. As of the September 2026 Help Center catalog, Pro subscribers get Sonar 2, GPT-5.6 Terra, Gemini 3.7 Flash, Claude Sonnet 5, Kimi K3, GLM 5.3, Grok 4.6, and Nemotron 3 Ultra. Max adds GPT-5.6 Sol and Claude Opus 5, and Enterprise tiers unlock all ten models on every plan. If your work depends on a specific frontier model — heavy coding, long-context analysis, creative generation — check which tier actually includes it before you subscribe. Perplexity's Help Center lays out the current per-plan catalog.

For buyers, the practical decision is matching tier to use case:

• Pro at $20/month covers the broad catalog and is enough for most search-and-synthesis work.
• Max matters if you need the top-tier reasoning models, like Claude Opus 5 or GPT-5.6 Sol.
• Enterprise unlocks all models for every seat, plus the strongest privacy posture — Enterprise data is never used to train AI models.
• Whatever you pick, treat the in-app selector as your source of truth, not third-party articles.

The deeper lesson from Perplexity is the one its own usage data makes vivid. In January 2025, just two models — Claude Sonnet 4 and GPT-4o — accounted for 91.5% of enterprise queries. By late 2025, no single model exceeded roughly 23% share, and 40% of power users actively juggle six or more models. Users stopped treating AI as one default tool and started treating it as a menu, matched to the task at hand.

That same "right tool per task" logic applies beyond model selection. If you're building AI agents, the skills you attach to them face the same problem Perplexity solved for models: too many options, no clear signal on which actually work. Published research found roughly 26% of agent skills carry at least one vulnerability, so choosing blindly is risky. SkillzBench is building a free, independent directory of agent skills — with version pinning, security scanning, and sandbox benchmarks that publish each skill's measured Lift — plus a free MCP connector so your agent only sees the skills you've chosen for it. You can join the waitlist now to get early access as the platform rolls out.

Match the model to the task. Match the skill to the job. Verify both.