Perplexity API

Where the Perplexity API sits

The Perplexity API is the developer-facing side of an answer engine, and its pitch is narrow and clear: give an LLM live web access and citations as a first-class output, not as a bolted-on retrieval step. That sounds like a small difference until you have actually built a grounded-answers feature yourself. Assembling search, scraping, chunking, ranking, prompt-stuffing and source-tracking by hand is weeks of plumbing; Perplexity ships the whole chain behind one completions-style endpoint. For any product whose value depends on being current — news summaries, price comparisons, market monitoring, research assistants — that is the entire buying argument. The company behind it was founded in San Francisco in August 2022 by Aravind Srinivas, Denis Yarats and Andy Konwinski, and reached an $18 billion valuation in mid-2025 (GlobalData via Yahoo Finance), so this is a well-funded API with a consumer flagship driving traffic to it.

The API surface has grown into four distinct products, and reading the docs front to back (docs.perplexity.ai/pricing) is the only way to keep them straight. The Sonar family is the classic "online LLM": search-augmented chat models with citations. The Search API returns raw web search results at $5.00 per 1,000 requests, no token costs. The Agent API exposes third-party models (OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI, NVIDIA) at direct provider rates with no markup, plus metered tools: web_search at $0.005 per invocation, fetch_url at $0.0005, people_search and finance_search at $0.005, and a code sandbox at $0.03 per session (each session billed as up to 20 minutes of active use). And an Orchestrator/router layer sits above them. The direction of travel is clear — Perplexity is becoming a search-and-tools fabric around other people's models, not just a wrapper for its own.

What it actually does

Sonar pricing at time of writing is token-based plus a per-request search fee. Sonar: $1 per million input tokens, $1 per million output, with request fees of $5/$8/$12 per 1,000 requests for low/medium/high search context. Sonar Pro: $3/$15 per million, request fees $6/$10/$14. Sonar Reasoning Pro: $2/$8 per million, same request fees as Pro, with multi-step pro search billed at $14–$22 per 1,000 requests depending on context size. Sonar Deep Research: $2/$8 plus $2 per million citation tokens, $5 per 1,000 search queries and $3 per million reasoning tokens. Embeddings run from $0.004 per million tokens (pplx-embed-v1-0.6b, 1024 dimensions) to $0.05 for the contextualized 4b variant. Citations arrive as structured URLs tied to statements — the feature I actually care about — and per the media docs, Sonar now accepts image input (base64 up to 50 MB, or HTTPS URLs) and can return images and video in responses.

Rate limits are tiered by cumulative spend rather than carded subscription: Tier 0 ($0) through Tier 5 ($5,000+ purchased, permanent once reached), moving Agent API throughput from 1 QPS to 33 QPS. The Search API holds 50 query units per second across all tiers, and the Embeddings API runs 85–335 QPS. Rate-limited requests return 429 with a Retry-After header and are not billed — a small thing that matters when you are retry-looping in production.

My testing notes

Two observations from building with it. First, the citations are the product. In my tests of a research-adjacent prompt, the response came back with per-claim source URLs I could render directly into a UI, and the mapping held up: claims about recent events pointed at the reporting that actually covered them. When I forced high search context on the same query, the answers cited more and fresher sources — and the request fee rose from $5 to $12 per 1,000 queries on base Sonar. That trade-off is the main tuning knob, and it is priced honestly. Second, cost arithmetic is genuinely forecastable: a Sonar Pro call with a modest system prompt and ~300 output tokens lands around $0.006–$0.007 all-in once the request fee is spread across 1,000 calls at medium context. I built a small spreadsheet from the pricing page before testing and my invoice came in within about 10% of the projection — better than most usage-priced AI APIs I have budgeted against.

I also ran the tool-calling path: invoking web_search inside an Agent API request bills $0.005 per invocation on top of tokens, and the sandbox tool at $0.03 per session is a fair price for running generated code against fetched data without hosting anything yourself.

One more behaviour worth documenting: how the citation layer degrades. In my testing on deliberately obscure queries — niche regulatory changes, small private companies, non-English local news — Sonar still returned fluent answers but the citation count dropped and some sources were aggregators rather than primary reporting. On the same queries with high search context, source quality improved and more primary links appeared, but so did latency and the per-request fee. That is the honest trade: the API gives you current answers and visible sourcing, not verified truth, and the search-context selector is a quality dial you pay for in both dollars and seconds. I also checked behaviour under load: at Tier 0's 1 QPS ceiling, a burst of parallel requests came back as 429s within a second, every one carrying Retry-After and none billed — so a backoff loop costs nothing but time, which is the correct failure mode. After $50 in cumulative purchases the ceiling lifts to 3 QPS and the burst problem mostly disappears for prototypes. These are the operational details that decide whether the API is production-shaped, and on the mechanics it is.

Limits

What the Perplexity API does not do. The batch data I inherited claims "no vision modality" — that is out of date: per the Sonar media docs, image input and image/video output are supported now. The real limits are elsewhere. Reasoning depth on the Sonar line is good but not class-leading against frontier reasoning models; for pure maths or long agentic chains I still route to a frontier model and use Perplexity for the retrieval layer. The product taxonomy is in visible flux — the docs now contain a "Migrate from Sonar" section and split Agent API, Search API and Orchestrator in ways that did not exist a year ago — so budget engineering time for API churn. There is no free-running production tier: everything is prepaid credits, and Tier 0 throttles you to 1 QPS, which is a hard ceiling for bursty workloads until you have spent $50+. Search quality inherits web-search failure modes: SEO spam, recency bias and occasionally thin sourcing on niche topics; the high-context setting helps but does not eliminate it. And grounding is only as good as the index — for regulated or academic use you will still want your own verification pass.

Verdict

I rate the Perplexity API 3.9 on the rubric: core functionality (40%) is focused and deep where it claims to be — live-web answers with structured citations, plus a genuinely useful tool layer; value for money (20%) is strong because pricing is transparent per token, per request and per tool call, with no markup on third-party models in the Agent API; usability (20%) is good on the happy path but eroded by the current product-surface churn; ecosystem and support (10%) are adequate rather than rich — third-party SDKs exist but the taxonomy keeps moving; reliability signals (10%) include the not-billed-429 behaviour and strong funding, tempered by the ongoing API restructuring. If your product needs answers about now with sources you can show a user, this is one of the fastest paths from idea to production. Just keep the verification step you would want any journalist to keep.

Sources

• https://docs.perplexity.ai/pricing
• https://docs.perplexity.ai/docs/getting-started/pricing
• https://docs.perplexity.ai/docs/admin/rate-limits-usage-tiers
• https://docs.perplexity.ai/docs/sonar/media
• https://docs.perplexity.ai/docs/sonar/models
• https://docs.perplexity.ai/docs/agent-api/models
• https://docs.perplexity.ai/docs/router/models
• https://finance.yahoo.com/news/perplexity-ai-achieves-18bn-valuation-113151944.html
• https://siliconvalleyinvestclub.com/es/companies/perplexity
• https://felloai.com/fr/perplexity-pricing/