How Perplexity Works:
Deep Research, Spaces, Pages,
Model Council, Comet, and More

Deep Research is the consumer feature, and sonar-deep-research is the API model that exposes the same capability programmatically. Both run an iterative agentic loop. The system decomposes the user query into sub-queries, performs parallel web searches, reads and evaluates source documents, updates its research plan based on findings, and iterates until a synthesis threshold is reached. The output is a structured cited report.

In the consumer interface, Deep Research queries take 2 to 4 minutes per execution. The user enters a query, selects Deep Research mode, and waits for the synthesis output. The report includes numbered citations linked to source URLs, can be exported to PDF or converted to a Perplexity Page, and contains roughly 30 searches per typical complex query.

The API exposes the same capability through sonar-deep-research with one important difference. The API has a 300-second timeout requirement that developers must configure. The default request timeout in many client libraries is shorter than this, and queries can fail silently if the timeout is not extended. The recommended pattern is to set request timeout to 300 seconds for any sonar-deep-research call.

The Variable Cost Structure

sonar-deep-research does not have a fixed per-query price. Total cost is the sum of five components.

• Input tokens at $2.00 per million.
• Output tokens at $8.00 per million.
• Citation tokens at $2.00 per million.
• Reasoning tokens at $3.00 per million.
• Search queries at $5.00 per 1,000 searches.

A single complex query (21 searches, 193,947 reasoning tokens, 19,028 citation tokens, 11,395 output tokens) costs approximately $0.82 per request based on official sample metadata. Cost projections built only on standard input and output token rates will significantly underestimate actual cost on long research queries. The reasoning tokens dominate cost on most complex queries.

Tier Availability

• Free tier: 5 Deep Research queries per day.
• Pro tier: approximately 500 Deep Research queries per day at average usage.
• Enterprise Pro: 50 per month.
• Enterprise Max: 500 per month.
• API: pay-per-use with the variable cost structure described above.

Documented Limitations

The HLE score for Deep Research at 21.1% (announced 2025-02-14) is now markedly stale. As of May 2026, the HLE leaderboard shows Gemini 3.1 Pro Preview at 44.7% and GPT-5.4 at 41.6% at the top. Perplexity has not published an updated HLE score for current Deep Research. The original benchmark claim is accurate at publication but the position has deteriorated significantly relative to the current frontier.

The CoT (chain-of-thought) tokens are not exposed in the API response for sonar-deep-research, unlike sonar-reasoning-pro which exposes reasoning in a <think> block. This is a deliberate design decision but limits debugging visibility for developers integrating Deep Research.

Sonar Reasoning Pro is the current premier reasoning model in the Sonar family, replacing the deprecated sonar-reasoning as of 2025-12-15. It uses enhanced multi-step chain-of-thought reasoning over real-time web search and outputs a <think> section containing reasoning tokens before the final response.

The exposed <think> block is a developer integration consideration. The response_format parameter does not strip these reasoning tokens, so developers requesting structured JSON output must implement custom parsers to extract the JSON portion of the response after the <think> block. This is a documented JSON parsing failure mode that affects integrations expecting clean structured output.

Sonar Reasoning Pro achieved a 1,143 score on the Search Arena leaderboard as of May 2026, ranking 11th globally with 29,825 votes. SimpleQA F-score is 0.858, the highest of any model at the time of testing per Suprmind’s AI Hallucination Rates and Benchmarks reference. GPQA Diamond is 62.3% per third-party leaderboard data, AIME 2025 is 77%, and MATH-500 is 92.1%.

The model uses 128K context window. Tool use and function calling are supported via the JSON output structure with the <think> prefix caveat. Cerebras inference is not used for the reasoning variants.

Spaces are workspace containers for related threads, files, and custom AI instructions. Users create named workspaces with optional custom instructions that apply to all threads within the space. Files can be uploaded directly (PDF, DOCX, and other formats) or pulled from connectors (SharePoint, OneDrive, Google Drive). The AI retrieves relevant sections from uploaded files at query time rather than loading the entire document into context.

Thread queries within a Space can toggle web search on or off and set recency filters. Custom AI instructions configured at the Space level apply to all threads inside that Space, so users can configure a “research assistant” Space with specific behavioral instructions and have those instructions apply to every conversation within it.

The persistence model differs from standard threads. Standard threads use a 7-day auto-purge for uploaded files. Spaces files persist until explicitly deleted. For workflows that require long-term reference document retention, Spaces is the structural fit.

Tier Availability

• Free: no Spaces access.
• Pro: up to 50 files per Space.
• Enterprise Pro: up to 500 files per Space.
• Enterprise Max: up to 5,000 files per Space.
• File size limits: 40 MB per file in consumer Spaces, 50 MB per file in Enterprise. Enterprise customers also get a 500-file organization repository.

Documented limitation: context window can fill with project files in long sessions, leaving limited space for conversation. Users running multi-thread Spaces with large file sets may hit context constraints inside individual threads even if the Space file count is below the tier cap.

Pages is Perplexity’s knowledge-creation feature. Users enter a topic and select an audience level (beginner or expert). Perplexity generates a multi-section article with inline citations and images sourced from current web data. The resulting Page is published on perplexity.ai with a shareable URL. Sections can be reordered, previewed, and unpublished by the creator. Pages can be added to Spaces.

The output format is a structured article with H1, H2, and H3 headings, image embeds, and inline numbered citations linking to source URLs. The article is automatically formatted for web reading and includes a publication URL for sharing.

Tier Availability and Limitations

Free tier provides basic Pages access (limited features). Pro tier provides full Pages including expert-level content generation, customizable sections, and media addition. Available via web and mobile.

Documented limitation: Pages cannot be exported to PDF, WordPress, or external CMS as of early 2026. This is the most cited Pages limitation in user feedback. For workflows that require export to external content management systems, Pages is structurally limited and the workflow either ends at perplexity.ai or requires manual content reconstruction in the destination CMS.

Model Council is Perplexity’s multi-model orchestration feature, launched 2026-02-05 and available exclusively at the Max tier ($200 per month) and Enterprise Max tier ($325 per seat per month). The feature dispatches a single user query to three frontier models simultaneously and produces a synthesis output.

The current configuration runs Claude Opus 4.6, GPT-5.2, and Gemini 3 Pro. The user submits one query. The query is sent to all three models in parallel. Each model produces an independent response. A separate synthesis model (the chair model) processes the three responses and produces a comparison output that explicitly surfaces points of agreement, points of disagreement, and unique insights from each model. The user sees both the individual model outputs and the synthesis.

Architectural Distinction

Worth flagging because the positioning overlaps with multi-model orchestration platforms. Model Council is parallel dispatch with synthesis: three models receive the same query independently, do not see each other’s responses, and the chair model summarizes after the fact.

Multi-model orchestration platforms run models in a shared conversation thread where each model reads what the others said before responding. The architectural difference produces measurably different outputs. Per the Suprmind Multi-Model Divergence Index, April 2026 Edition (n=1,324 production turns), 99.1% of multi-model turns produce at least one contradiction, correction, or unique insight that single-model use would miss. Shared-thread orchestration captures cross-model corrections embedded in the response sequence rather than reported in a separate synthesis layer.

Tier Availability and Limits

Max ($200 per month) and Enterprise Max ($325 per seat per month) only. Not available on Pro. Plans announced for future Pro tier expansion but not yet implemented as of the research date. Web only at launch. Three models fixed in current configuration. Users cannot select which three models participate. The $200 per month Max tier creates a high cost barrier for professional users who want multi-model validation.

Perplexity Labs is a self-supervised work cycle that runs approximately 10 minutes per task. Labs combines web browsing, code execution, chart and image generation, and asset creation to deliver structured outputs. Unlike Deep Research, which primarily synthesizes sources into a report, Labs executes code and creates interactive or exportable files.

Output deliverables include reports, spreadsheets, dashboards, and web apps. Labs positions closer to an agentic project assistant than a research tool. The user describes a project objective, Labs runs the work cycle, and the output is a set of files or a deployed asset.

Tier Availability

Pro, Max, Enterprise Pro, Enterprise Max. Free tier does not have Labs access. The “Create files and apps” query quota differentiates tiers: Pro gets monthly average-use limits, Enterprise Max gets 500 per month.

Documented limitation: 10-minute maximum self-supervised work cycle per task is the documented hard limit. Output format constraints are not publicly detailed beyond the general structured-deliverable category. Labs is relatively new and user feedback is still developing.

Comet is Perplexity’s AI-native desktop browser for Mac and Windows, built on Chromium with a sidecar AI assistant embedded in every tab. The assistant can answer questions about the current page, summarize content, perform cross-tab tasks, manage email, handle shopping, and execute background agentic tasks.

Comet launched 2025-07-09 as Max-only and was made free for all users worldwide on 2025-10-01. The browser download is free. Feature access within Comet is gated by Perplexity subscription tier, with Free users getting limited daily queries and Max users getting full access including the background “Comet Assistant” agentic capability.

Tier Availability

Comet browser download is free for all users worldwide. Feature access within Comet is gated by Perplexity subscription tier. Comet Plus is a $5 per month add-on for the publisher content bundle, available to all Comet users including Free.

Perplexity Shopping interprets purchase intent from conversational queries, retrieves live product data (pricing, availability, specifications, reviews) from integrations with Shopify, Amazon, BigCommerce, and other marketplaces, and presents curated product cards.

Three core capabilities differentiate Shopping from standard web search. “Snap to Shop” allows image uploads to find visually similar products. “Instant Buy” is a checkout button (built with PayPal, supporting 5,000+ merchants) that enables in-session checkout without leaving Perplexity. “Buy with Pro” is a direct purchase mechanism for supported merchants available to Pro subscribers.

The launch timeline: late 2024 saw the initial “Buy with Pro” button. November 2025 expanded the feature with free AI shopping for all US users.

Tier Availability and Geographic Scope

Free for all US users (web and desktop) as of 2025-11. Instant Buy and checkout features require Pro. US-only initially. Amazon checkout redirects to Amazon rather than completing in Perplexity.

The Citation System is core to Perplexity’s product positioning and was built into the original product rather than launched separately. At generation time, Perplexity’s search index retrieves candidate documents. The LLM generates a response and attaches numbered inline citation markers ([1], [2], etc.) to claims in the response. The API response includes a citations array of URLs and a search_results array with title, URL, date, and snippet for each cited source.

Sonar Pro delivers approximately 2x more citations than standard Sonar. The search_context_size parameter (Low, Medium, High) controls how much retrieved web evidence is injected per turn, which affects citation density and accuracy.

CJR Audit Result

The Columbia Journalism Review’s Tow Center tested eight AI search platforms in 2025-03 on news article citation tasks. Perplexity Sonar Pro answered 37% of queries incorrectly, the lowest error rate among tested platforms. ChatGPT Search: 67%. Grok 3: 94%. The 37% rate is the best in the field but still substantial in absolute terms.

A separate measurement on the “Pro variant” specifically reported 45% error rate per the Suprmind AI Hallucination Rates and Benchmarks reference. The variant-level distinction matters because the Pro variant is the model most users on the Pro subscription receive in standard usage.

The Structural Failure Mode

Per Suprmind’s AI Hallucination Rates and Benchmarks reference (May 2026 update), Perplexity’s hallucination pattern is structurally distinct from non-citation hallucination. The model cites real URLs with content that may be fabricated. The URL is genuine. The claim attributed to it may be invented. This is harder to detect than non-citation hallucination because the URL creates an appearance of verifiability that the user does not have time to manually audit.

The system also does not distinguish between citations that came from parametric training knowledge versus claims grounded via live web retrieval within a single response. All citations are nominally from the search results array, but the relationship between the model’s knowledge sources and the final cited URLs is not exposed at the per-claim level.

Citation Source Distribution

Independent research suggests Perplexity prioritizes Reddit at 46.7% of citations in one study. The source distribution affects citation quality: Reddit content quality varies enormously, and the visual treatment of Reddit citations alongside primary sources or peer-reviewed material does not consistently differentiate source authority.

Perplexity accepts a broad set of file formats across consumer and API surfaces. Files in standard threads are auto-purged after a tier-specific retention period. Files in Spaces persist until explicitly deleted.

Supported Formats

• Documents: PDF, DOC, DOCX, TXT, RTF, ODT.
• Spreadsheets: XLSX, CSV (25 MB recommended maximum for best parsing results).
• Presentations: PPTX.
• Text and code: Markdown, JSON, HTML.
• Images: PNG, JPEG (vision analysis).
• Audio: MP3, WAV, OGG, FLAC (up to 40 MB, transcribed to text).
• Video: MP4, MOV, AVI, WEBM (up to 40 MB, transcription only).

The audio and video processing is transcription-based rather than full multimodal understanding. The model receives the transcript text rather than processing the raw audio or video stream natively.

Plan-Based File Limits

Parser fidelity: XLSX and CSV parsing is recommended at 25 MB or less for best extraction accuracy. Larger files may produce degraded extraction. OCR for scanned documents was listed as “coming next” in September 2025 documentation, so scanned PDF fidelity may remain limited as of the research date. For workflows that depend on scanned PDF extraction, test empirically before relying on the parser output.

Tier availability for several features is not enumerated in official Perplexity documentation. Treat tier-specific limits as Volatile and verify at perplexity.ai before relying on the cap for production planning.