Sonar Deep Research: Budgeting a Report, Not a Reply

Most model calls return an answer in seconds and cost a fraction of a cent. Sonar Deep Research does neither. One request kicks off a research job: many searches, a lot of reading, a long reasoning pass, then a report with sources. Treat it like a chat completion and both your timeout and your cost estimate will be wrong.

The model has been listed on OpenRouter since March 2025, a few weeks after Perplexity launched Deep Research in its app. This guide is about operating it: how the bill is built, how to estimate it before you run a job, how to handle calls that take minutes, and why Perplexity's own documentation now points new work elsewhere.

What one request actually does

When Perplexity announced Deep Research in February 2025, it described the product as performing dozens of searches, reading hundreds of sources and reasoning over the material to produce a comprehensive report, typically in two to four minutes. The API model follows the same pattern. OpenRouter describes it as autonomously searching, reading and evaluating sources, and refining its approach as it gathers information.

OpenRouter lists a 128,000-token context window and text-only input. You send a research question, not a document; the model finds its own material.

Five meters on one request

Perplexity's pricing page bills Sonar Deep Research on five separate dimensions:

• Input tokens: $2 per million.
• Output tokens: $8 per million.
• Citation tokens: $2 per million, for the source material it reads.
• Reasoning tokens: $3 per million.
• Search queries: $5 per thousand searches.

OpenRouter lists the same input and output rates, $3 per million for internal reasoning, and $0.005 per search at the time of writing. The input you send is the smallest line. The model decides how many searches to run and how much to read, so the cost of a request is set by the question, not by your prompt length.

Estimating a report before you run it

There is no fixed per-report price, so estimate from assumptions and then replace them with measurements. A purely illustrative job:

• 30 searches at $5 per thousand: $0.15
• 150,000 citation tokens at $2 per million: $0.30
• 50,000 reasoning tokens at $3 per million: $0.15
• 8,000 output tokens at $8 per million: about $0.06

That lands around 66 cents, with the prompt contributing almost nothing. The real figures depend on the question, and broad questions read more. Run ten representative jobs, record each meter from the usage block, and budget on the 90th percentile rather than the average. Agent cost control patterns covers why the tail, not the mean, is what surprises finance.

Perplexity added a reasoning_effort parameter for this model in 2025, which adjusts how much computation each query gets. It is the main lever you have: lower effort for scoping questions, higher for the final pass.

Calls that take minutes

A two-to-four-minute request breaks assumptions all over a typical stack. Default HTTP client timeouts, serverless function limits, reverse proxies and load balancers often cut connections well before that. Streaming keeps the connection alive, but a dropped stream four minutes in loses the work you have already paid for.

In 2025 Perplexity offered an asynchronous API for Deep Research: submit, then poll. Its current documentation says asynchronous Sonar requests are no longer supported and points users to background mode on its Agent API instead. If you built on the old async flow, it needs migrating. If you are starting fresh, put the call in a background job with its own generous timeout and store the result, rather than holding a user request open. Agent timeout strategies covers how to set those bounds.

The launch numbers, read in context

At launch Perplexity reported 21.1% on Humanity's Last Exam and 93.9% on SimpleQA for Deep Research in its app. Those were February 2025 figures for the consumer product, on benchmarks that have moved on considerably since, so they say little about how the API model compares with today's research agents.

A better test is your own: ten questions where you already know the right answer and the right sources. Score whether the report reaches the right conclusion, and check a sample of citations by hand, because a report that cites real pages for claims those pages do not make is the characteristic failure of this kind of system. How to benchmark LLMs on your own work has the method.

Perplexity now points elsewhere

Perplexity says Sonar Chat Completions support ended on 27 September 2026, with synchronous and streaming requests being reformulated as Agent API requests model by model. Its migration guide maps Sonar Deep Research to the Agent API high preset, describes that preset as often costing less per request than Sonar Deep Research, and recommends the xhigh preset for the most demanding research.

When a vendor says its replacement is often cheaper per request, take it seriously, while noting that its quality claims rest on internal benchmarks. If you call Perplexity directly, evaluate the presets before committing more work to this model. If you reach it through OpenRouter, where the listing continues, keep a fixed question set and rerun it periodically to catch behaviour changes.

When to use it

• Use it for scheduled or background research where a sourced, multi-page report is the deliverable: market scans, literature overviews, due-diligence drafts a person will review.
• Do not use it for interactive answers. For a precise answer to a hard question, Sonar Pro Search is faster; for a single fact, Sonar is far cheaper.
• Always keep a human review step before a report is published or acted on.