Perplexity combines search with AI-generated answers, making it useful for exploring a topic and following its sources. Sonar became known as Perplexity’s search-grounded API offering, with different models intended for quick answers, deeper search and reasoning. The attraction is an answer alongside evidence you can inspect, rather than having to assemble every initial lead yourself.
You can also read our Perplexity AI article. This guide explains the Sonar concept, the current API transition and a practical way to judge the output. Citations make checking easier, but they do not establish that every generated statement is correct.
What Sonar means
Sonar is not simply a universal website button that reveals the complete internal thought process behind every answer. The Sonar name has been used for Perplexity’s search-grounded model/API family, including reasoning-oriented variants. Search results, citations, an explanation and internal computation are different things.
A generated explanation can help a reader understand a conclusion. A citation can lead to a document that supports a factual claim. Neither guarantees that the explanation is a faithful transcript of hidden computation, or that the cited page supports every nearby sentence. Evaluate the evidence, not the apparent confidence of the wording.
The current API transition
Perplexity’s migration documentation states that Sonar Chat Completions support ended on 27 September 2026. It recommends the Agent API for new projects and migrating existing Sonar integrations.
The same documentation says synchronous and streaming Sonar requests continue through a gradual reformulation into Agent API requests. Asynchronous Sonar requests are no longer supported through that arrangement and need the documented background-mode migration. This is a more precise distinction than assuming every old endpoint has stopped responding or every old request remains supported unchanged.
The older Sonar site link remains useful as a reference to the product’s history. For new integration work, follow the current API console and documentation. Do not assume an older screenshot, model list or third-party tutorial describes the present interface.
From model names to presets
The current migration guide suggests these starting points. They describe a route for migration, not identical architectures or guaranteed equivalent answers:
| Earlier Sonar choice | Suggested Agent API starting preset |
|---|---|
| sonar | fast |
| sonar-pro | fast |
| sonar-reasoning-pro | low |
| sonar-deep-research | high |
Perplexity manages the underlying configuration of a named preset. Its model, prompt, tools and parameters can change as the service evolves. If a reproducible application requires a fixed configuration, use the documented configuration or profile approach and retain the settings you selected. A preset name alone does not permanently identify one foundation model.
Model architectures, context limits and supported parameters can change as the service evolves. Check the supported model catalogue and parameters for the exact request you plan to make. Select a workload and compare quality, delay and cost on representative questions.
A small current API example
The Agent API quickstart documents POST requests to /v1/agent. This Bash/cURL example uses a managed fast preset and a non-sensitive technical question. It requires an API account, a valid key and a cURL version supporting the shown options; running it can incur service charges.
# Bash example; set a valid API key in the environment before running.
: "${PERPLEXITY_API_KEY:?Set PERPLEXITY_API_KEY first}"
curl --fail-with-body --silent --show-error \
https://api.perplexity.ai/v1/agent \
-H "Authorization: Bearer $PERPLEXITY_API_KEY" \
-H "Content-Type: application/json" \
--data '{
"preset": "fast",
"input": "Explain what IEEE 802.11ax means and distinguish Wi-Fi 6 from Wi-Fi 6E. Cite primary technical sources and state any regional limitations."
}'
Supply the key through a suitable environment or secret-management mechanism. Keep it out of source repositories, shared transcripts and client-side browser code. The command prints the raw response so you can inspect its structure before writing a parser. Check the HTTP status and structured response before adding it to an automated workflow.
The Agent API uses an input and typed output items, rather than the old messages/choices layout. Source lists appear in search_results output items. An integration should collect the results from every relevant item and map citation identifiers to those structured results. Do not try to construct authoritative links solely from URLs the model spells out in answer text.
Checking an answer
- Match the subject: confirm that the source describes the exact product, country, version or period you asked about. A related page can still answer a different question.
- Open the evidence: read the supporting section, including qualification notes. A search snippet may omit the conditions that make a specification meaningful.
- Separate fact from recommendation: an official port specification is different from someone’s opinion that a router is good value.
- Check time: distinguish publication date, revision date and the date of the event. A recent crawl does not make an old announcement current.
- Allow uncertainty: if the answer has no direct supporting source, treat it as unresolved and refine the search rather than filling the gap with confidence.
For example, ask about the ports of a specific regional router revision, then compare the answer with that revision’s specification page. If two sources disagree, record the difference and check whether they cover different hardware. More citations are not necessarily better if all point to copies of one underlying claim.
Prompting for useful results
Start with a focused question, the relevant context and the output you need. “Compare these two mesh systems for a home with wired backhaul and a 1Gbps connection” is more useful than “Which router is best?” Add exact model names and a country when specifications or availability may differ.
Ask for the assumptions and limitations behind a recommendation, and request that missing evidence be identified. For API integrations, use supported search parameters when you need enforceable source, date or region constraints. A sentence asking for only one domain is not equivalent to configuring an actual retrieval filter.
A follow-up question should target the uncertainty left by the previous answer. Ask which source establishes an unsupported feature, or whether a quoted maximum is a radio link rate rather than measured application throughput. This turns the answer into a starting point for research instead of an unexamined final verdict.
Research, learning and comparisons
Search-grounded answers are useful for getting oriented, finding primary documentation and outlining a comparison. In education, ask for a worked explanation and then check the supporting definitions. For fact-checking, trace the precise claim to the underlying source and look for contrary evidence rather than relying on the tool’s label alone.
Our ChatGPT reasoning article covers another approach to AI assistance. Compare services by the actual capabilities available in your chosen mode: source retrieval, citations, response quality, supported tools, privacy and cost. Visible explanatory prose by itself is not a sound ranking of transparency or truthfulness.
Cost, privacy and a practical verdict
Review current API billing before adding a request to a recurring workflow. Model usage and enabled tools can affect charges, so measure representative runs, limit unnecessary repetition and handle errors deliberately. Test parsing and missing-source behaviour as well as successful answers before relying on an automated report.
Keep passwords, customer records and private infrastructure details out of casual research prompts. Review the relevant service terms and your own handling requirements before sending sensitive material. A public source citation does not establish that an input was processed locally or never retained.
Sonar’s appeal remains the convenience of well-organised answers with evidence to explore. The current Agent API is the documented route for new integrations, while careful prompting and source checking remain valuable whichever interface you use. Use the service to accelerate the first stage of research, then verify the facts that determine your decision.
