Perplexity Free vs Pro vs Max: what to choose

Free is a useful baseline for ordinary search with citations and occasional harder queries. Pro is worth testing when work regularly needs expanded Pro Search, Research, manual selection of advanced models, file analysis, or artifact creation. Max is not simply about “better answers”; it targets power users who need the highest consumer access, Max-only capabilities, or materially more capacity on supported surfaces. No paid tier removes the need to verify primary sources.

Start with a concrete bottleneck. If Free passes your acceptance criteria, you do not need a subscription. If Pro delivers acceptable research outcomes without critical interruptions, Max is likely unnecessary spend. If you need organization ownership, seat management, SSO, retention policy, audit, or centralized data controls, do not buy personal Max as a workaround: evaluate Enterprise Pro or Enterprise Max. The API also has separate usage and billing contracts and is not included in a consumer subscription.

• Occasional search and source verification → start with Free.
• Regular Research, model choice, and file analysis → test Pro.
• A proven capacity bottleneck or Max-only workflow → run a controlled Max trial.
• Team rights, retention, and lifecycle → move to Enterprise evaluation.
• Programmatic integration → evaluate API, credits, and budget controls separately.

The official plan guide separates Search, Pro Search, Research, browser-agent queries, advanced models, and other surfaces, but availability and limits can change. The model catalog explicitly calls the selector in your account the source of truth and warns about plan, mode, region, staged rollout, and organization settings. Do not preserve a feature table as an eternal fact. At decision time, record account type, country, billing channel, plan, renewal cadence, visible modes, model selector, quota wording, Computer credits, Comet access, and the source link.

Eligibility is not capacity, and capacity is not outcome. A visible Research mode only proves you can launch it; it does not guarantee enough runs or an acceptable report. A third-party model name inside Perplexity also does not mean equivalence with its native app: search orchestration, system instructions, tools, citations, and limits belong to Perplexity. Mark unknown values as `unknown`; do not reconstruct quota from an old review or somebody else’s screenshot.

Free currently offers practically unlimited basic search according to the official description, but only limited access to Pro Search, Research, and files, and no manual choice of advanced search model. That is enough to validate the core loop: formulate the question, open citations, find the supporting passage, inspect date and scope, and separate fact from synthesis. Create ten real fixtures: a fresh fact, conflicting sources, an academic query, a product comparison, a local-language request, a PDF, a table, a paywall, a missing page, and a request where the correct answer is to abstain.

Pro makes sense when expanded searches, Research, model selection, uploads, or Create files and apps regularly become accepted outcomes. Do not count only response length or citation volume: verify passage entailment, freshness, source diversity, conflict disclosure, coverage, and reviewer time. More links can mean broader evidence, but they do not prove every claim is supported. If the bottleneck is a weak question or no authoritative source exists, an upgrade will not repair the workflow.

• Freeze → identical questions, source requirements, and acceptance criteria.
• Capture → mode, model, citations, elapsed review time, and limit event.
• Verify → consequential claims in an opened primary source.
• Compare → accepted outcomes, not query count or report polish.
• Upgrade → only when Pro removes a recurring cause of failure.

Perplexity’s official Max description positions it as the top consumer tier with everything in Pro, the highest access to advanced models, enhanced Research and Create files and apps, Max Assistant in Comet, and early access to new capabilities. Individual features such as Model Council on web and iOS can be Max-only, while Computer uses its own credits. These are different purchase drivers: model eligibility, mode eligibility, capacity, browser execution, and metered agent work cannot be reduced to one word, “Max”.

Run a crossover on Pro and Max with unchanged fixtures. For Research, compare coverage and citation audit; for Create, artifact correctness and reproducibility; for Comet, exact target, confirmation, permission boundary, and postcondition; for Computer, accepted task per credit spent and behavior after timeout. Max passes the gate only if a required capability is genuinely unavailable on Pro or measured interruptions cost more than the upgrade. Vendor language such as highest access is an entitlement claim, not an independent accuracy or performance metric.

Perplexity starts as an answer engine with citations, but Comet, Create files and apps, and Computer expand the surface into artifact creation and actions. For the search gate, require a supporting passage, correct revision, coverage of must-have questions, conflict disclosure, and no leakage. For the artifact gate, add formulas, schema, file integrity, and reproducible inputs. For a browser or connected-tool action, require exact identity, permitted resource, operation, material parameters, confirmation, idempotency key, receipt, and rollback owner.

Negative tests matter more than a demo. Add prompt injection in a retrieved page, a similarly named wrong destination, a revoked connector, a target change after approval, and a timeout after a possible write. A correct workflow stops before an unknown or broader action and reconciles the system of record before retry. A higher tier is not permission for greater authority: a capability can be technically available and still prohibited by your policy.

• Search PASS → the claim is supported by the exact passage in a current source.
• Artifact PASS → the file is correct, opens, and can be reproduced from inputs.
• Action PASS → exact target, approval, receipt, postcondition, and rollback.
• Unknown scope or ambiguous destination → fail closed.
• Surface, connector, or role change → targeted regression before reuse.

Current Perplexity documentation says AI Data Retention is enabled by default for Free, Pro, and Max, while an individual user can disable the use of future data for AI training in settings; opting out does not remove previously collected training data and does not stop processing needed to operate the service, provide security, or meet legal obligations. Verify the actual toggle, change date, incognito behavior, search history, Memory, Brain, upload retention, and connected tools. Do not project Enterprise policy onto a personal plan.

Enterprise documentation separately describes no-training treatment, admin controls, roles, retention, seat management, and identity integrations, with some controls depending on contract and rollout. If work contains confidential company data or must survive an employee departure, personal Pro or Max creates an ownership gap regardless of answer quality. Store the approved report, source ledger, decision, and hash in a managed system of record, not only in a personal session or Project.

In week one, run 12–20 frozen tasks on the current plan; in week two, run them on the candidate tier. Count accepted reports, unsupported material claims, correction minutes, blocked runs, credits, subscription, tax, and review cost. For sensitive decisions, set zero tolerance for critical citation or authority failure. Cost per accepted outcome is more useful than cost per query: a large allowance has no value if nobody verifies the results or the workload does not use the paid capabilities.

Before upgrading, record the billing channel: web, iOS, and Android can have different subscription-management paths, and the official Max guidance warns about the risk of a separate mobile subscription when moving from web Pro. Before downgrading, export required records and inspect Projects, files, Memory/Brain, connectors, credits, and renewal state. Cancel the subscription first, then decide separately about deleting the account: current guidance explicitly warns that deleting the account by itself does not cancel the subscription.

• Baseline → current plan, fixtures, receipt, and failure taxonomy.
• Candidate → the same workload without specially improved prompts.
• TCO → subscription + credits + review + failures + migration.
• Decision → the lowest tier that passes the critical gates.
• Exit → cancel, export, revoke, verify billing, and set a review date.