Perplexity Token Counter & API Pricing (2026)

Perplexity charges twice for the same query. Once for the tokens. Once for the search. Miss the second line and your cost estimate will be off by 2x or 3x on short questions, because the request fee on Sonar, Sonar Pro, and Sonar Reasoning Pro is billed per 1,000 requests, not per token. Every generic token calculator built for OpenAI or Anthropic rates will under-read a real Perplexity bill. This guide walks through every published Sonar rate, the per-request fee, Sonar Deep Research's multi-meter math, and the Sonar Chat Completions sunset set for September 27, 2026.

Where Perplexity Stands in 2026

Perplexity is the only major provider that ships a model and a live web search in the same API call. OpenAI, Anthropic, and Google all require you to bolt search on top via tools or external services. Perplexity bundles it. The trade-off is billing complexity.

The pricing docs give the formula directly: total cost per query = token costs + request fee. The request fee scales with how much web content the call pulls, not with your prompt length. Nothing you type into Sonar changes the fee — what changes it is how hard Sonar has to search.

There is a second quirk the docs spell out. The tokens Perplexity bills you for are not only the tokens you sent. Retrieved web pages enter the working context, and the output carries citations and reference material. Perplexity's own glossary lists "search results and references" under output tokens. A 12-token prompt can produce a 700-token answer that silently carries grounded source text — and you pay the output rate for all of it.

A third difference matters for anyone planning past late 2026. Sonar Chat Completions is moving to the Agent API, and the docs state Sonar is supported until September 27, 2026. After that date, the Agent API becomes the only path. The Agent API resells third-party models — OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI, NVIDIA — at direct provider rates with no markup, plus per-tool charges. So building on Sonar today is a short-to-mid-term commitment. Building on the Agent API is the long-term play.

Sonar Models & 2026 Pricing

All four Sonar models published on docs.perplexity.ai. Prices in USD per million tokens unless noted.

Sonar Deep Research is the only model with citation, reasoning, and search-query meters. The other three charge tokens plus a flat request fee.

Per-Request Search Fee

The request fee is what most people forget. It is charged per 1,000 successful requests, varies by search context size (low / medium / high), and is the same for Sonar, Sonar Pro, and Sonar Reasoning Pro.

Low is the default and the cheapest. High gives maximum search depth. Switching search context size does not change token rates. Pro Search on Sonar Pro bumps the request fee to $14 / $18 / $22 per 1,000 while keeping token rates at $3 and $15.

The math consequence: on a 200-token-in / 500-token-out Sonar query at low context, the search fee is 88% of the total bill. Doubling the prompt length barely moves the total. Doubling the search context triples the request fee.

Two adjacent products are worth knowing. The Search API is flat — $5 per 1,000 successful requests, no token charges, and a request with up to five queries counts as one billing unit. Embeddings run from $0.004 per million tokens on pplx-embed-v1-0.6b up to $0.05 per million on the contextualized 4B model. Agent API tools are priced per call: web_search $0.0025, fetch_url $0.0005, people_search and finance_search $0.005 each, sandbox $0.03 per session.

Sonar Deep Research Multi-Meter Breakdown

Sonar Deep Research is the only model with four independent meters. You cannot predict how many searches it will run. The docs say the model decides, and the reasoning_effort parameter only nudges it.

Worked example for one report: 2,000 input + 6,000 output + 3,000 reasoning + 1,500 citation + 20 searches.

• Input: 2,000 / 1,000,000 × $2 = $0.004
• Output: 6,000 / 1,000,000 × $8 = $0.048
• Reasoning: 3,000 / 1,000,000 × $3 = $0.009
• Citation: 1,500 / 1,000,000 × $2 = $0.003
• Search: 20 / 1,000 × $5 = $0.100

Total: $0.164 per report. The search-query meter is 61% of it. This is the most common source of surprise invoices in the Sonar lineup. Teams running Deep Research at scale should monitor usage.search_queries in the API response, not just token counts.

Sonar Chat Completions Sunset (2026-09-27)

Perplexity has moved Sonar Chat Completions to the new Agent API. The Sonar Chat Completions endpoint is supported until September 27, 2026, per the official pricing docs. After that date the endpoint stops accepting new requests.

The migration path:

• Today (Sonar Chat Completions): bundled model + search. One bill, token rates + request fee.
• After sunset (Agent API): model + tools as separate line items. Token rates from the underlying provider (OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI, NVIDIA), plus per-call tool fees (web_search, fetch_url, etc.). Perplexity takes no markup on the model tokens.

Two practical consequences:

1. The Sonar-specific request fee disappears. You pay the underlying model's tokens plus per-tool invocations. A 500-in / 200-out call on openai/gpt-5.2 with one web search on the Agent API is roughly $0.000875 + $0.0028 + $0.0025 = $0.006175 (this is the docs' own worked example).
2. The Sonar house models (Sonar, Sonar Pro, Sonar Reasoning Pro, Sonar Deep Research) are still available on the Agent API as named endpoints, but their pricing shape changes — request fees become tool fees.

Plan the move now if you depend on Sonar. The pricing model shifts from bundled to à la carte on the same calendar date the endpoint retires.

Real Bill Examples

All math below is computed at published rates, not estimated. Rounded to four decimal places.

Example 1 — Single Sonar query, low search context. 200 input, 500 output. - Input: 200 / 1,000,000 × $1 = $0.0002 - Output: 500 / 1,000,000 × $1 = $0.0005 - Request fee: $5 / 1,000 = $0.005 - Total: $0.0057

The request fee is 88% of the bill. Tokens barely register.

Example 2 — Single Sonar Pro query, medium search context. 500 input, 900 output. - Input: 500 / 1,000,000 × $3 = $0.0015 - Output: 900 / 1,000,000 × $15 = $0.0135 - Request fee: $10 / 1,000 = $0.01 - Total: $0.025

Now output tokens are the largest line — 54% of the total. The $15/M Sonar Pro output rate is what flips the cost shape.

Example 3 — 10,000 Sonar Pro queries per month, medium context. Using Example 2's token shape: 10,000 × $0.025 = $250/month. Switching to Sonar at low context with the same token counts: 10,000 × ($0.0005 input + $0.0009 output + $0.005 request) = $64/month. A 74% cut just from picking the right model and search tier.

Example 4 — Search API instead of Sonar. 10,000 search-only requests, no LLM in the loop. 10 × $5 = $50. Bundle up to five queries per request and the effective per-query cost drops to $0.001. If you only need results and are running your own model, this is the cheapest path.

Frequently Asked Questions

How much does the Perplexity API cost per token? Sonar is $1 per million in and $1 per million out. Sonar Pro is $3 and $15. Sonar Reasoning Pro is $2 and $8. Sonar Deep Research is $2 and $8 with three extra meters (citation $2, reasoning $3, search $5/1K). Add the per-request fee on the first three.

Does the Perplexity Pro subscription include API tokens? Not in any meaningful way. Perplexity Pro is the consumer product at $20/month with a small monthly API credit (around $5, reported in third-party 2026 pricing roundups). API usage beyond that bills from prepaid credits at the published per-token rates. The subscription and the API are two separate wallets.

Is Perplexity cheaper than adding web search to another model? Often yes at the Sonar tier, because grounding is bundled into one bill. At the Sonar Pro tier the $15 output rate erases most of that advantage. Run Example 3 with your own token counts before deciding.

What is a citation token? A token spent producing search results and references — source links, reference numbers, bibliographic detail. It is metered separately and only on Sonar Deep Research, at $2 per million. You cannot predict how many citation tokens a report will generate; log them from usage.citation_tokens in the API response.

What happens after September 27, 2026? Sonar Chat Completions stops accepting new requests. The Agent API becomes the only path. Pricing shape changes from bundled search (token + request fee) to à la carte (underlying model tokens + per-tool fees). Perplexity takes no markup on the underlying model tokens.

Rates verified September 7, 2026. Perplexity changes pricing without notice — check docs.perplexity.ai before budgeting. Nothing here is financial advice.