Perplexity Sonar Guide: Cheap Search Answers and Hidden Fees

The usual way to compare models is price per million tokens. For Sonar that number is close to useless. It is the cheapest token rate among the Sonar models, and for most real workloads the tokens are the smaller part of what you pay.

Sonar is Perplexity's lightweight search-grounded model: you send a question, it searches the web, and it answers with the sources it used. It has been listed on OpenRouter since January 2025, and in 2026 the API it lives on started changing underneath it. Both points matter more than any benchmark if you are deciding whether to build on it.

What Sonar is

When Perplexity introduced its new Sonar model in February 2025, it said the model was built on Llama 3.3 70B and further trained for factuality and readability. Perplexity's own migration guide describes the intended job as short: single-fact lookups, definitions and quick summaries, with citations attached.

The useful property is that retrieval is built in. You do not run a search API, fetch pages, chunk them and stuff them into a prompt; one call does the lot and returns results alongside the answer. In 2025 Perplexity replaced the old citations field with search_results, so read sources from there. OpenRouter lists Sonar with text and image input and a context window of 127,072 tokens.

If you are weighing built-in search against your own retrieval stack, LLM-powered search explains why the retriever, not the model, usually decides answer quality.

The request fee is most of the bill

Perplexity prices Sonar at $1 per million input tokens and $1 per million output. On top of that it charges a request fee per thousand calls: $5 at low search context, $8 at medium and $12 at high. OpenRouter lists the same token rates plus $0.005 per request for search at the time of writing.

Run the arithmetic on a typical lookup workload. Ten thousand questions at around 700 tokens each, prompt plus answer, is 7 million tokens, or about $7. The request fee on the same ten thousand calls is $50 at low context and $120 at high. Tokens are a rounding error next to it.

Two consequences follow. First, shortening prompts saves almost nothing; avoiding unnecessary calls saves almost everything. Cache answers to repeated questions and do not send a search model questions that need no search. Second, comparing Sonar with a plain chat model on token price alone understates its cost by a factor of several. Check the usage block on a few real responses before you budget, because what is billed is not always what you typed.

Search context size is the real dial

The low, medium and high setting controls how much web material Perplexity retrieves for each request. Token prices do not change with it; only the request fee does. More context means more sources to ground the answer, which helps on questions where the first page of results is thin or contradictory.

For the job Sonar is meant for, low is usually enough. A definition or a single fact rarely improves with more sources. Raise it per request for the questions where you can show it helps, rather than globally. A simple rule: start at low, log every answer a user flags as wrong, and rerun those at high. If high fixes most of them, route that question type to high.

The name is changing underneath you

Perplexity's changelog for July 2026 records that Sonar Chat Completions is now its Agent API, and its documentation says Sonar Chat Completions support ended on 27 September 2026. Synchronous and streaming requests keep working and are being reformulated as Agent API requests, rolled out gradually by model. Asynchronous Sonar requests are no longer supported.

Perplexity's migration guide maps Sonar to the Agent API fast preset, which its presets page lists as a single-step configuration using Fast Search and an OpenAI model. The docs do not say whether reformulated Sonar requests run on that preset's model, and they say nothing about third-party routes such as OpenRouter.

The practical reading: behaviour behind the name sonar may shift without the name changing. If answer style or accuracy matters to you, keep a small set of fixed questions and rerun them weekly. For new projects on Perplexity directly, its docs recommend the Agent API, and that is where future changes will land first.

Sonar or Sonar Pro

Sonar Pro costs $3 input and $15 output per million tokens, with request fees of $6, $10 and $14 per thousand. It is aimed at harder, multi-step questions and takes a larger context. Notably, Perplexity's migration guide maps both Sonar and Sonar Pro to the same fast preset, which suggests the gap between them is narrower than the price gap implies.

Pick Sonar for high-volume lookups where each answer is one fact. Pick Sonar Pro when questions need several sources reconciled and you have measured that Sonar gets them wrong. For questions that need many searches and a report, neither is right: look at Sonar Deep Research.

A five-minute evaluation

1. Write 25 questions your users actually ask, with answers you can verify.
2. Run them at low and high search context.
3. Score correctness, and separately whether the cited source supports the claim. A right answer with a wrong citation is a liability in any product that shows sources.
4. Compute cost per correct answer, including the request fee.

If low and high score the same, you have just found the cheapest configuration. If they do not, you know which questions are worth paying for.