The Perplexity API is a set of developer products, mainly the Orchestrator API, the Agent API and the Search API, and each is billed differently. As of October 2026 the model APIs are billed per token at each model's published rates, tools are charged per call on top, and the Search API is charged per request, with no model and no tokens involved.
The catch: you pay per request, and empty results still count
Most people assume a search API charges per query. Perplexity's Search API doesn't. It charges per successful POST /search request, and one successful request can carry an array of up to five queries. Send five questions in one request and you've used one billing unit. Send them as five separate requests and you've used five.
The other half of the catch is less pleasant. A successful request that comes back with no results is still billed. What isn't billed is an invalid request, a rate-limited request, or one where something failed upstream. So a search that's too narrow, with tight domain, language and date filters, can cost you without returning anything useful.
If you're building something that runs lots of searches, batch your queries five at a time and test your filters on a small sample before you let the job loose.
The three APIs and what each one bills for
- Orchestrator API. Billed per token at each model's published rates, with no per-request fees. Every model has its own input and output rate, cache reads are billed at a discounted rate, and reasoning tokens are billed at the output rate. You're billed at the rates of the model you asked for, however Perplexity serves the request.
- Agent API. Gives you third-party models from OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI and NVIDIA, priced per token at each model's published rates. Tool calls are charged separately from the tokens.
- Search API. Returns ranked web results as JSON, with page text extracted into each snippet. There's no model, no generated answer and no token billing. You can filter by domain, language and date, and bias results by country.
The model token rates sit on Perplexity's rate card for each model, so I won't repeat them here; they change by model. What matters for your budget is knowing that the Agent API bill has two parts, tokens and tools, and the second part is easy to forget.
Tool and search prices as of October 2026
These are the per-call and per-request prices from Perplexity's pricing docs.
| What you call | Price | What it does |
|---|---|---|
| Search API | $5.00 per 1,000 requests | Ranked web results as JSON, no model |
| Search API with Fast Search | $1.00 per 1,000 requests | Same format, lower latency |
| web_search (Agent API tool) | $0.0025 per call, or $0.001 with Fast Search | Searches the web inside an agent run |
| image_search | $0.0025 per successful call | Image results with filters |
| fetch_url | $0.0005 per call | Pulls content from a URL |
| people_search and finance_search | $0.005 per call each | Professional profiles and market data |
| sandbox | $0.03 per session | Code environment, 20-minute billing window |
Fast Search is the obvious saving on both the Search API and the web_search tool. If your app doesn't need the slower standard search, try Fast Search first and only switch back if the results aren't good enough.
How the sandbox is billed
The sandbox is the one most people misread. A session covers up to 20 minutes of active use for billing purposes. Perplexity's docs are clear that this is a billing window, not a runtime cap, so your code won't be cut off at 20 minutes. The window is how they count sessions.
There's a second line item hiding inside it. Search queries made from inside the sandbox are billed per query at the standard web search rate, on top of the session charge. An agent that writes code, runs it and searches the web from the same sandbox will show both on the bill.
Working out what an agent run will cost
An Agent API call can add up four separate charges. Here's how I'd estimate one before you put it into production.
- Count the input and output tokens for your chosen model, plus any reasoning tokens, which are charged at the output rate.
- Add one charge for each tool call the agent makes: web_search, image_search, fetch_url, people_search or finance_search.
- Add a sandbox session charge for each session the agent opens.
- Add the per-query search charge for any searches run inside the sandbox.
Run a small batch of real requests, then compare your estimate with the usage figures in the Admin section, where billing setup and usage monitoring live. Rate limits and usage tiers have their own page in the docs, so check those before you scale up.
Who the Perplexity API suits
If you only want answers for yourself, you don't need the API at all. My Perplexity AI review covers the app, and this guide to using Perplexity shows you how to get the most from it. The API is for people building something: a developer adding live web search to an app, or a small business owner wiring search into a no-code tool.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
The Search API is the simplest fit for that second group. You send queries, you get ranked results with the page text already pulled out, and you pay per request with no tokens to count. If you need the AI to write an answer from those results, that's where the Agent API and its token bill come in.
My verdict on the Perplexity API
I'd use the Search API if you need fresh web results inside your own tool, because the pricing is easy to predict once you batch five queries per request. I'd use the Agent API if you want a choice of models from several providers through one bill, as long as you budget for tool calls on top of tokens. I'd skip it if you're just after better answers for your own work; the app does that without a developer account. And remember that an empty result still costs you, so test your filters before you run anything at scale.
Curious about the neighbouring question? Try my Perplexity vs Grok breakdown.
Related: erp: guidelines and how to pitch.
Frequently asked questions
How is the Perplexity Search API billed?
Per successful POST /search request, not per query. As of October 2026 one request can hold up to five queries and still counts as one billing unit.
Do I pay for a Perplexity search that returns no results?
Yes. A successful request with no results is billed. Invalid, rate-limited and upstream-failed requests aren't.
Does the Perplexity Search API use tokens?
No. It returns ranked web results as JSON with no model, no generated answer and no token billing.
Which models can I use through the Perplexity Agent API?
Third-party models from OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI and NVIDIA, each at its own published token rate.
Is the Perplexity sandbox limited to 20 minutes?
No. The 20 minutes is a billing window for each session, not a runtime cap.
