How to Use Perplexity AI Models: A Practical Walkthrough

Why Perplexity Is Now an Agent Platform, Not Just a Search Engine

Choosing the right AI tool for real work has never felt harder — every week brings a new model, a new API, and a new promise that this one will finally fit your workflow. Perplexity's evolution is worth understanding precisely because it stopped being one more chatbot and became something more useful: a platform for routing, researching, and building agents.

The shift is architectural, not cosmetic. As one analysis puts it, Perplexity is now "a router over the best models plus live web search, not a single-model chatbot" — model choice becomes a per-query decision rather than a lock-in. Its lineup spans in-house models (Sonar, R1 1776) alongside frontier third-party models like GPT-5.x, Claude, Gemini, and Grok. A third-party developer guide describes the platform as having "evolved from an AI-powered search engine into a full developer platform," anchored by four APIs:

• Search API — single-pass web search with citations, built on the Sonar model
• Agent API — iterative retrieval, tool calling, and context continuity via sonar-pro
• Embeddings API — using sonar-embedding for hybrid RAG pipelines
• Sandbox API — an upcoming addition for running agent-generated code

The business signal behind this pivot is hard to ignore. Perplexity's reported ARR growth — from roughly $100M in early 2025 to over $450M by March 2026, per Financial Times figures cited in the source — is attributed to the shift toward AI agents and usage-based pricing. Reported monthly query volumes have grown into the hundreds of millions, though figures vary across sources and dates, so treat them as directional.

What agents are actually used for matters even more. Perplexity's field study of "hundreds of millions of interactions" found that about 57% of agent activity is cognitive work — analysis, research, decision support — with agents acting as "a thinking partner rather than a simple butler." That reframes the practical question: not "which chatbot do I subscribe to," but "which platform lets me build research-and-analysis workflows?"

One caveat keeps this honest: independent testing by Columbia Journalism Review's Tow Center found even the best AI search tool had a 37% citation error rate, so verification steps belong in every workflow. For teams extending agents further with skills and custom tools, SkillzBench is building a free directory and MCP connector that security-scans and benchmarks agent skills before you equip them — a sensible guardrail when published research found roughly 26% of skills carry at least one vulnerability.

Choosing the Right Model and API for Each Query

Model choice in Perplexity is not a one-time commitment. The platform functions as a router over the best models plus live web search, turning model selection into a per-query decision rather than a lock-in. That framing changes everything about how you build agent workflows.

The cleanest rule of thumb comes from a third-party developer guide: if a query can be answered with a single web search, use the Search API. If it requires search, evaluate, and search again, use the Agent API. The Agent API's three differentiators are iterative retrieval (multiple web searches per request), tool calling (custom functions alongside web search), and context continuity where each search builds on prior results.

Two config options matter for most production pipelines:

• search_recency_filter (hour/day/week/month) controls how fresh the results need to be
• return_citations should stay on; citations double as a trust signal for users and a debugging tool for developers

Before committing production traffic to any single model, the Model Council feature runs three frontier models in parallel and compares outputs. This is the practical way to A/B test responses and benchmark quality before switching default traffic, rather than guessing which model handles your domain best.

Deep Research mode is worth the wait when accuracy matters more than latency. It runs on Opus 4.6 and reportedly achieves state-of-the-art performance on the Google DeepMind Deep Search QA benchmark. For routine lookups it is overkill; for multi-source synthesis where a wrong answer costs more than a few extra seconds, it pays for itself.

One caveat that should shape every workflow: independent testing by the Tow Center found a 37% citation error rate even for the best-performing AI search tool, with more than 60% of responses across eight tools containing citation errors. Treat citations as starting points for verification, not substitutes for opening the source. Vendor-reported accuracy figures are directional at best, and the source-linked format is the real safety net.

For teams extending agent capabilities, custom tool calling alongside web search is where the leverage compounds. Agent skills (SKILL.md instruction packages) can automate adjacent steps in research and analysis workflows, and SkillzBench is building a free directory and MCP connector to discover, verify, and benchmark those skills with security scanning included, relevant given that roughly a quarter of published skills carry at least one vulnerability.

Building the Think-Act-Observe Workflow Step by Step

Perplexity's field study of hundreds of millions of interactions found that 57% of agent activity is cognitive work — agents acting as "a thinking partner rather than a simple butler," per the study's analysis. That finding should shape how you build agent workflows: structure tasks around reasoning, not chores.

The study defines agents as systems that cycle automatically between three iterative phases: thinking, acting, and observing. To implement this with Perplexity, start by deciding which API handles each phase. A developer guide to the platform offers a simple rule: if a query needs a single web search, use the Search API (sonar); if it requires search, evaluate, then search again, use the Agent API (sonar-pro).

The Agent API's three differentiators map directly onto the loop:

• Iterative retrieval — multiple web searches per request, powering the think and act phases.
• Tool calling — custom functions running alongside web search, so agents can act on your systems.
• Context continuity — each search builds on previous results, closing the observe phase.

Turn citations on permanently in production by setting return_citations. The guide's advice is blunt: citations serve as both a trust signal for users and a debugging tool for developers. That matters because independent testing by Columbia Journalism Review's Tow Center found a 37% citation error rate even for the best-performing AI search tool, and more than 60% of responses across eight tools contained citation errors, per that analysis.

Handle rate limits deliberately. The Search API and Agent API maintain separate rate limit pools, so heavy search traffic won't starve your agent calls. When you hit a 429, read the Retry-After header and apply exponential backoff rather than hammering the endpoint. A 401 means your key is invalid — fix that before debugging anything else.

Finally, delegate subtasks, not whole jobs. The field study found power users make 9x as many agentic queries as average users, breaking work into smaller pieces and handing off only pieces of it. If you're extending agents with custom tool calling, agent skills (SKILL.md instruction packages) can automate the surrounding workflow — SkillzBench is building a free directory and MCP connector to discover, verify, and benchmark them, which matters given research finding roughly 26% of published skills carry at least one vulnerability.

Verifying Agent Output Before You Act on It

Citations look authoritative. They're also, disturbingly often, wrong — and the better the tool, the easier it is to forget that.

Independent testing by Columbia Journalism Review's Tow Center found that even the best-performing AI search tool had a 37% citation error rate, and more than 60% of responses across the eight tools tested contained at least one citation error, with rates ranging up to 94%. Perplexity topped the list — which means the winner still got its sources wrong roughly one time in three. That should reframe how you read every answer an agent hands you.

The Tow Center's findings cut deeper than a single stat. As that analysis notes, premium chatbot subscriptions produced more confidently incorrect answers than their free counterparts — paying more doesn't buy accuracy, just more polished-sounding mistakes. The habit to break, per the same source, is "treating any AI answer engine's citation as a substitute for opening the actual source."

That makes verification a human-in-the-loop discipline, not a nice-to-have. A practical check before you act on any agent output:

• Open the cited source directly and confirm it actually says what the agent claims — the citation existing is not the same as the citation supporting the answer.
• Check that the source is current and authoritative, not a stale or low-quality page the retrieval step happened to grab.
• Cross-check any load-bearing fact against a second, independently retrieved source before it enters a deliverable.
• Treat vendor accuracy claims as directional, not independently verified — as one analysis puts it, the source-linked format is the feature you can actually rely on, because it lets you click and check.

This skepticism should extend to how you extend the agent itself. Perplexity's Agent API supports custom tool calling alongside web search, which is where SKILL.md instruction packages come in — but published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building a free, independent answer: a directory that security-scans every skill, pins each version to a commit SHA, license-checks it, and benchmarks it in a sandbox with and without the skill to measure the real Lift. Its free MCP connector will let you equip any agent with only the verified skills you choose, so extending your Perplexity-style workflow doesn't mean importing someone else's unvetted code.

The pattern is the same at every layer: agent output, cited sources, and the skills that power the agent all earn trust through verification, not through confidence. Build the checking step into the workflow, and the speed of agentic search stops being a liability.