A Complete Guide to Today’s Perplexity AI Models (2026 Edition)
If you’ve opened Perplexity recently and found yourself staring at a model selector populated with names from OpenAI, Anthropic, Google, Moonshot AI, xAI, Z.ai, and NVIDIA all at once, you’re not alone. The platform has evolved well beyond a simple AI chatbot, and understanding what you’re actually choosing – and why – requires a different mental model than most AI tools.
This guide covers the current Perplexity lineup, how the platform actually works, and how to make sensible choices without turning every search into a decision tree.
Perplexity Isn’t a Chatbot. That’s the Point.
The most useful framing for Perplexity isn’t as a rival to ChatGPT. It’s as an AI-powered research environment – one where web search, source retrieval, citations, and AI reasoning are integrated into a single workflow rather than separated across different tools.
That distinction shapes everything about how you should think about model selection inside Perplexity.
When you choose Claude, GPT, or Gemini inside Perplexity’s interface, you’re not getting an identical experience to what you’d find in those companies’ standalone products. Every third-party model runs inside Perplexity’s own search infrastructure, with its own prompts, safety systems, tooling, and usage limits layered on top. The same underlying model can and does behave differently here than it does natively.
This isn’t a flaw – it’s the product. Perplexity’s value is the layer it builds around those models: real-time web retrieval, source citations, research workflows, and the ability to switch between AI providers without leaving the platform. The model is one component of a larger system.
As of August 2026, Perplexity Pro and Max subscribers can choose from Sonar 2, GPT-5.6 Terra, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Sonnet 5, Claude Opus 5, Kimi K3, GLM 5.2, Grok 4.5, and Nemotron 3 Ultra — depending on their subscription tier. The lineup changes frequently and without notice, so your account’s model selector is always the authoritative source.
The Current Model Lineup
Sonar 2: Perplexity’s Native Option
Sonar is the one model Perplexity built itself, and it remains the most natural starting point for users who want the platform’s core search experience without selecting a third-party model.
It’s available to both Pro and Max subscribers and doesn’t offer a Thinking toggle. Its strength is in what Perplexity does best: finding current information, answering factual questions with citations, synthesizing web content quickly, and handling research tasks where retrieval quality matters more than extended reasoning depth. For everyday searches – looking something up, getting a quick summary, pulling together current information on a topic – Sonar 2 is the native choice.
GPT-5.6 Terra: OpenAI’s Balanced Option
OpenAI’s GPT-5.6 Terra is the more widely accessible of the two OpenAI models in Perplexity’s lineup, available to both Pro and Max subscribers with optional Thinking.
OpenAI positions Terra as the balanced member of the GPT-5.6 family – below the flagship Sol in raw capability, but offering strong general-purpose performance across analysis, coding, structured writing, and knowledge work. Inside Perplexity’s search environment, it’s a reliable choice when you want an OpenAI model’s reasoning paired with real-time web retrieval. For demanding tasks where more deliberate reasoning is useful, Thinking can be enabled.
GPT-5.6 Sol: OpenAI’s Flagship, Max Only
Sol is OpenAI’s most capable model in the current Perplexity lineup and is reserved for Max subscribers.
OpenAI describes it as the flagship of the GPT-5.6 family, with particular strength in coding, knowledge work, scientific reasoning, browsing, and tool use. Within Perplexity, it’s the right call when you need OpenAI’s strongest model for genuinely demanding work – complex analysis, professional research, or tasks where the extra capability tier is worth the access cost. For routine searches, it’s more than you need.
Claude Sonnet 5: Anthropic’s General-Purpose Workhorse
Claude Sonnet 5 is available to both Pro and Max subscribers, with optional Thinking. It represents Anthropic’s current general-purpose flagship – the model Anthropic describes as its most agentic Sonnet yet, with meaningful improvements in reasoning, tool use, coding, and knowledge work over its predecessors.
Inside Perplexity, it’s a strong default for software development questions, coding and debugging, technical analysis, multi-step knowledge work, and tasks that benefit from structured reasoning. One thing worth noting for users familiar with older Perplexity interfaces: Sonnet no longer appears as a separate “Thinking” model in the selector. Thinking is now a toggle on Sonnet 5 itself, which you enable when the task warrants it.
Claude Opus 5: Anthropic’s High-End Option, Max Only
Opus 5 is Anthropic’s premium offering inside Perplexity, available exclusively to Max subscribers. It occupies the high end of Anthropic’s lineup when you need more capability than Sonnet provides – difficult reasoning, complex professional analysis, nuanced or lengthy tasks, strategic work, and situations where the additional reasoning depth justifies the higher cost of access.
For most everyday questions, Sonnet 5 will handle the job. Opus becomes the better choice as task complexity and stakes increase.
Gemini 3.1 Pro: Google’s Always-On Reasoning Model
Gemini 3.1 Pro is available to both Pro and Max subscribers, with one notable characteristic: Thinking is always enabled, not optional.
Google launched Gemini 3.1 Pro in early 2026 as a reasoning-focused model designed for complex problem-solving – situations where a direct answer without deliberation isn’t sufficient. Inside Perplexity, that makes it a natural choice for technical and scientific questions, data-heavy research, coding tasks that benefit from deeper analysis, and any context where you want reasoning to be a default part of every response rather than an optional mode to switch on.
Kimi K3: Moonshot AI
Kimi K3, from Moonshot AI, is available to Pro and Max subscribers with Thinking always enabled. Its inclusion reflects how significantly Perplexity’s model lineup has expanded beyond the major US AI labs.
For users running parallel comparisons – asking the same question across different models to identify where they converge and where they diverge – Kimi K3 adds a meaningfully different voice to the comparison. Enterprise users should note that Kimi K3 is not currently listed in Perplexity’s Enterprise Search model table, so consumer availability doesn’t carry over to managed workspaces.
GLM 5.2: Z.ai
GLM 5.2 from Z.ai is another Pro and Max option with Thinking always enabled. Like Kimi K3, its presence signals Perplexity’s intent to offer a genuinely diverse range of AI providers rather than a US-centric lineup.
It’s worth experimenting with for reasoning-heavy tasks or as a second opinion on complex questions. It is similarly not listed in Perplexity’s Enterprise model table.
Grok 4.5: xAI
Grok 4.5 from xAI is available to both Pro and Max subscribers with optional Thinking. It functions as another general reasoning option within Perplexity’s search environment, and the most practical use case is the same as with the other alternative models: comparing its output against other models on important questions, or using it as a standalone choice when you prefer xAI’s approach.
Nemotron 3 Ultra: NVIDIA
Nemotron 3 Ultra is NVIDIA’s current model in Perplexity Search, available to both Pro and Max subscribers with Thinking always enabled. It replaces the earlier Nemotron 3 Super. For analytical and technical questions where you want another always-on reasoning option, it rounds out a lineup that now spans providers across the US, China, and beyond.
Understanding “Thinking”
One of the most frequently misunderstood aspects of the current Perplexity lineup is how Thinking works – partly because older versions of the interface presented it differently.
Thinking is not a separate model. For models that support it, Thinking is a mode that instructs the model to reason more deliberately before producing its response. Some models have it available as an optional toggle; others have it permanently enabled; Sonar 2 doesn’t support it at all.
The practical implication is simple: for complex, multi-step, or ambiguous questions, enabling Thinking (where available) can meaningfully improve output quality. For fast factual lookups, it’s likely unnecessary overhead.
The Modes That Matter as Much as the Models
Model selection is only part of how you work in Perplexity. The platform also offers distinct modes that change what you’re actually asking the system to do.
Best: The Default Worth Using
Perplexity’s own recommendation for most users is to leave the model set to Best and let the platform choose automatically. This is sound advice. For everyday questions, the overhead of manually selecting a model rarely produces better results than Perplexity’s own routing logic – and it removes friction from the majority of searches.
Manual model selection earns its place when you have a specific preference, you’re doing technical work where a particular model’s strengths are relevant, or you want to compare outputs across providers.
Deep Research: When Depth Is the Point
For serious research tasks, the more consequential choice isn’t which model to select – it’s whether to use Deep Research.
Deep Research conducts repeated web searches, reads numerous sources, reasons through what it finds, and synthesizes the material into a comprehensive report. Perplexity says the system can perform dozens of searches and read hundreds of sources in a single research task. Recent updates have also added the ability to run calculations, work with a code sandbox, analyze uploaded documents, and access harder-to-reach corners of the web.
The use cases are broad: market research, competitive analysis, due diligence, technology assessments, academic research, complex product comparisons, and detailed professional reports. Any task where a single web search would produce a thin or incomplete answer is a candidate for Deep Research.
One important distinction: in Deep Research mode, you don’t manually select the underlying model. Perplexity handles model selection automatically as part of the research process. That makes Deep Research a distinct workflow rather than simply another entry in the model selector – and it’s worth treating it that way.
A related note on terminology: Perplexity’s API includes a model called sonar-deep-research for developers, with its own separate access and billing. That’s a developer tool, not a consumer model option. In the consumer application, Deep Research is a mode, not a model you select.
Model Council: Multiple Models on the Same Question
Model Council is available to Max and Enterprise Max subscribers and does something no individual model can: it sends the same query to three models simultaneously and synthesizes their responses into a unified answer – while surfacing where the models agree and where they diverge.
Users can adjust which models participate and enable Thinking for individual models where supported.
The value here is structural. For high-stakes questions – investment decisions, market research, complicated strategic choices – the point of disagreement between models can be as informative as any individual answer. Rather than manually running parallel queries and cross-referencing the results, Model Council turns multi-model comparison into a built-in workflow.
Pro vs. Max: What Access You Actually Get
The split between Perplexity’s paid consumer tiers is primarily about which models and modes you can access.
Pro subscribers can currently reach Sonar 2, GPT-5.6 Terra, Gemini 3.1 Pro, Claude Sonnet 5, Kimi K3, GLM 5.2, Grok 4.5, and Nemotron 3 Ultra. Max subscribers get those models plus GPT-5.6 Sol, Claude Opus 5, and access to Model Council.
Enterprise availability is a separate matter. Administrators can restrict which models appear in a workspace, and the Enterprise model table differs from the consumer lineup in meaningful ways. Organizations building on Perplexity should verify their specific workspace configuration rather than assuming the consumer lineup applies.
A Practical Starting Point for Each Use Case
Rather than an objective ranking – which doesn’t exist in a meaningful way across tasks this varied – here’s a reasonable starting point by use case:
For everyday questions, start with Best. For general web search and fast retrieval, Sonar 2 is the native choice. For coding and technical work, Claude Sonnet 5 and GPT-5.6 Terra both perform well. For demanding reasoning or analysis, Gemini 3.1 Pro (always-on Thinking), GPT-5.6 Sol (Max only), or Claude Opus 5 (Max only) are the natural escalation points. For alternative model perspectives, Kimi K3, GLM 5.2, Grok 4.5, and Nemotron 3 Ultra each bring a different approach to the same prompt. For comprehensive research, Deep Research is the right mode entirely. For cross-checking conclusions, Model Council.
These are starting points, not rankings. Model performance is prompt-dependent, providers update their systems frequently, and the lineup itself changes without announcement. Perplexity explicitly acknowledges this – which is why they recommend checking the model selector in your account for the current state rather than relying on any external guide, including this one.
The Bigger Picture
The most useful thing to understand about Perplexity is what it actually is: not a bet on one AI model, but a platform built around the idea that the model is one component of a larger research system. The web search layer, the retrieval infrastructure, the citation tooling, the research workflows — those are what differentiate a Perplexity answer from the same model answering the same question on its own.
For businesses that work with information –researching markets, tracking competitors, synthesising technical documentation, generating analysis for clients – that integrated layer is genuinely valuable. The question isn’t really which model is best. It’s whether the research environment around the model is suited to the kind of work you’re doing.
For most of those tasks, the answer is yes. And for most of those searches, Best is still a fine place to start.