What Model Does Perplexity Use? Sonar and Model Choice

Perplexity uses its own Sonar model alongside third-party models, with seven selectable Search models listed for Pro. The practical answer to “what model does Perplexity use” depends on your mode: Best chooses automatically, paid Search lets you choose, and Research selects its own combination of models.

What Model Does Perplexity Use by Default?

Sonar is Perplexity's own model for search, but a default setting labeled Best does not guarantee that Sonar answers. That distinction matters when you need to know which model produced a recommendation.

Perplexity's default-model guidance describes a model optimized for quick searches and web browsing, available to free and Pro users. That page does not name it. The company's Sonar announcement identifies its in-house model as trained for default search; newer model-selection guidance describes Best as automatic selection for each query.

The defensible answer is therefore more precise than “everything runs on Sonar.” Sonar explains Perplexity's own model work. Best describes how the app can choose an answering model.

Verified against Perplexity's public Help Center and its Sonar post on November 8, 2026. The model table below retains the vendor's own October 27 snapshot date. This is a documentation check, with no signed-in testing.

Perplexity AI Model Name: Model, Mode and Product

Perplexity is the product; Sonar is a model; Search and Deep Research are ways to use the product. A large language model, or LLM, is the system that interprets language and generates an answer. The surrounding product also decides how to gather evidence and present it.

Think of a research desk: the sources are the documents on the desk, the mode is the assignment, and the model is the analyst drafting the answer. Changing the analyst can change the explanation. Changing the assignment from a quick question to a research report changes the work being requested.

For an operator comparing suppliers, this distinction prevents a common mistake: trying another model when the answer needs a broader research assignment or a better primary source.

Does Perplexity Have Its Own LLM?

Yes. Perplexity calls Sonar its in-house model, and describes further training an existing base model for its search experience. “Its own” does not mean the announced model was built entirely from scratch.

In its Meet New Sonar announcement, Perplexity says that version was built on Llama 3.3 70B and further trained to improve factuality and readability. Those are useful goals for research: interpreting retrieved evidence and presenting it clearly.

That statement identifies the base of the Sonar version described in that post. It does not establish the architecture of every later Sonar deployment, nor which model handled your latest Best response. Avoid treating a historical architecture statement as a live execution log.

Does Perplexity Have Its Own Model?

Yes, Sonar. Perplexity also integrates models from other providers into its own search experience. These arrangements coexist: having an in-house model does not mean every answer must use it.

For a founder deciding whether the product fits a research workflow, the useful question is whether its answers provide the evidence and explanation the decision needs. Ownership of the underlying model alone does not settle that.

Which Models Can Each Perplexity Plan Select?

Pro provides manual Search model selection; Max adds another group of models. Free users can use Best and limited Pro Searches, but cannot manually select an advanced Search model.

The official advanced-model availability page supplies the following plan mapping. Its tables are dated October 27, 2026, and were checked on November 8. Availability can change; your account's model selector is the final check.

The providers are OpenAI for GPT, Google for Gemini, Anthropic for Claude, Moonshot AI for Kimi, Z.ai for GLM, xAI for Grok, and NVIDIA for Nemotron. The counts above are our count of the vendor's listed rows, not usage allowances.

This table applies to Search. Computer, Comet, Model Council and the API have their own selection or availability rules. Sonar's absence from this particular advanced-model table is also a reason to avoid presenting it as a guaranteed manual choice in every current account.

For a mid-market CTO, the Enterprise caveat matters more than a public model list. Administrators can restrict providers or individual models; regional requirements and data-retention settings can also affect availability. Trials, promotions and staged rollouts can make an account's selector differ from the standard tables.

For subscription costs, use our Perplexity pricing guide.

What Do Pro Search and Research Use?

Pro Search can use your eligible Search model choice; Research chooses its models automatically. Neither label is the name of a single LLM.

Pro Search Works Within Search

Perplexity's Pro Search instructions say there is no separate Pro Search toggle. Choose Search and ask your question; the product applies Pro Search automatically when your plan includes it, up to the relevant limit.

It performs multiple searches, synthesizes source material and provides citations. Leave the model on Best, or choose an eligible model. The newer advanced-model table above is the source for current selections; the Pro Search article still contains older model examples.

For a senior operator checking a supplier's published integration support, begin with Search and ask for the official documentation. The first acceptance criterion is whether the cited page supports the claim.

Research Chooses a Combination of Models

The Research mode guidance explicitly says that it automatically selects a combination of models and does not let you manually choose a specific one. Free access is limited, while Pro provides extended access.

There is a narrower, dated model note: the Advanced Deep Research update, last modified September 15, lists Opus 4.6 Thinking for Max and “4.5 Thinking” rolling out gradually to Pro. The Pro label is incomplete on that page, so it should not be expanded into a guessed model name.

Those are the update's stated models, not an exhaustive current roster for every Research task. The newer general model page confirms automatic selection without publishing a complete version-by-version routing map. Selecting a model in Search therefore does not prove that Research will use it.

For a founder commissioning a market-entry brief with competing evidence and unresolved questions, choose Research for the report-shaped task. Do not try to force it into a preferred Search model.

How to Choose a Model and Check Which One Answered

Check the answer after it arrives, because the model selected before sending is not always the model that responds. Perplexity documents fallback when a model is unavailable, unsuitable for the request, or restricted by a usage limit. There is no setting that prevents this substitution.

The following procedure comes from its model-selection and answer-identification instructions.

1. 1

   Choose an eligible Search model

   On the web, open the model dropdown at the bottom right of the prompt box. Select Custom, then choose the model and the available reasoning and speed options. Leave Best selected when you want automatic selection.

2. 2

   Inspect the answer's model

   Open the three-dot menu below the answer, then choose Rewrite. The model that answered is preselected in the model list.

3. 3

   Keep the uncertainty visible

   If you used Best, or the answering model is absent from the list, the menu shows Best. Record it that way instead of assigning a model name you cannot verify.

4. 4

   Compare only when the answer needs it

   To request another answer, choose a different model from the list and select Rewrite. Compare the evidence, unsupported assumptions and usefulness of the explanation, not just its tone.

The Best Perplexity Model Depends on the Work

Start with Best for routine questions; change the setting when you can name what the first answer failed to do. This is a workflow recommendation, not a benchmark ranking of the listed models.

For a senior operator preparing a supplier shortlist, an unsupported claim calls for better evidence. Ask for a primary source and inspect it. Another fluent answer is not a substitute for a source that establishes the fact.

For a solo technical builder untangling a question with several dependencies, try Thinking where the selected model supports it. Perplexity describes Thinking as extra reasoning time, with a possible latency trade-off. It is a setting on a model, not another model to add to the plan count.

For a funded founder requesting a broad competitive report, move to Research when the assignment requires sustained investigation. Accept that this changes who controls model selection.

For a mid-market CTO documenting a repeatable workflow, keep the prompt, date, mode, requested model and available answer attribution with the sources. If exact model identity is mandatory, an answer labeled only Best does not meet that requirement.

Wait before upgrading solely for a longer model list. First identify a recurring task that the current selection does not handle well, then compare eligible alternatives against the same acceptance criteria. For the broader product decision, read our Perplexity review.

What's Overhyped About Model Choice?

Access to a model inside Perplexity does not reproduce the provider's standalone app. Perplexity combines it with its own search, citations, tools, prompts, safety systems and usage limits. A familiar model name does not make those surrounding systems identical.

The other overclaim is treating a model list as a quality ranking. The official availability table establishes access, not which model produces the best answer for your business question. Nor does it make a citation self-validating: open the source and check whether it supports the sentence attached to it.

The largest practical limitation is attribution. A manual selection can fall back, and Best may conceal the exact answering model. For audit-sensitive work, preserve that uncertainty instead of converting a preference into a claim of reproducibility.

Frequently Asked Questions

What model does Perplexity AI use?

It uses its own Sonar model and third-party models. In Search, Best chooses automatically and paid plans can select eligible models. Research automatically chooses a combination of models.

Is Perplexity an LLM model?

Perplexity is the product that combines search with language models. Sonar is its in-house model; the product also offers models from other providers.

Are Perplexity and ChatGPT the same thing?

No. Access to OpenAI models inside Perplexity does not make it the same product as ChatGPT. Perplexity supplies its own search, citations, tools and usage limits.

Is Perplexity more accurate than ChatGPT?

The cited Help Center pages do not establish a universal accuracy winner. Judge a research answer by whether its sources support its claims, and compare products on the same task before making that decision.

What are the disadvantages of Perplexity AI?

For model control, the documented limits are fallback to another model, incomplete model identification under Best, and automatic selection in Research. A subscription also does not include API usage; the API has a separate catalog and billing.

Does Perplexity AI track you?

Do not infer privacy from the selected model. The model guidance says consumer Pro and Max users can opt out of AI training data collection, while Enterprise data is not used to train AI models. That addresses training, not every form of tracking or data collection.

How to calculate Perplexity of language model?

This asks about a language-model evaluation metric rather than which model powers the Perplexity product. The product documentation cited here does not provide a calculation method.

Is Perplexity making a profit?

The Help Center and Sonar announcement used here do not establish company profitability. Model availability is not evidence of profit.

What are the models of AI?

For this product, distinguish Perplexity's Sonar model from the third-party Search models listed above. A taxonomy of AI models beyond Perplexity is outside these sources.

What are the 5 things you shouldn't tell ChatGPT?

These Perplexity sources do not establish a five-item rule for ChatGPT. For workplace research, follow your organization's data-sharing requirements rather than treating access to a model as permission to submit information.

Your Next Research Session

Choose the assignment first, then the model setting. Take a recurring work question, state what evidence a useful answer must include, and start with Search on Best. Use an eligible alternative or Thinking if the explanation needs improvement; choose Research if the task needs a full report. Before sharing the result, check the supporting sources and record whatever model attribution the interface reveals.

Subscribe to the newsletter for practical guides to using AI in your work.