Sonar Deep Research
Language Search-grounded Deprecated Full comparison ↗ perplexity/sonar-deep-research · by Perplexity
· decoder-only-transformer
Search-grounded model. Unlike a standard base model, this offering combines an underlying LLM with live web retrieval and citations — not a directly comparable like-for-like with a non-search-grounded entry. Billed with the token pricing below plus a per-request search-context fee (see Pricing).
Pricing — 1 offering(s)
About this offering: Perplexity's deep-research model. Unlike the other Sonar models it has no search-context-size tiers; instead it bills five things separately: input, output, citation and reasoning tokens, plus a fee per search query. Perplexity ended support for Sonar models on Chat Completions on 2026-09-27, but it still serves those requests while it moves each model to its Agent API. Its pricing page no longer shows the Sonar tables, so the prices here are the last published rates. Asynchronous Sonar requests no longer work at all; Perplexity maps Sonar Deep Research to its Agent API "high" preset ("xhigh" for its best deep research).
Input tokens
- $2.00 / 1M tokens (input) Current 2026-07-26 → present
Output tokens
- $8.00 / 1M tokens (output) Current 2026-07-26 → present
Citation tokens
- $2.00 / 1M tokens (citation) Current 2026-07-26 → present
Reasoning tokens
- $3.00 / 1M tokens (reasoning) Current 2026-07-26 → present
Search queries
- $5.00 / 1k requests Current 2026-07-26 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Sonar Deep Research fits into a cost-aware routing setup
See how →Capability profile
How Sonar Deep Research rates across core capability dimensions, with the task-level evidence behind each rating.
Perplexity's "expert-level research model" — designed for multi-step retrieval, synthesis, and reasoning across complex topics, handling "hundreds of sources for detailed reporting."
The defining capability: autonomous, iterative web search — it issues many search queries per task (billed at $5 / 1,000 queries) and adapts the search plan as it reads. Registry entry sets `agentic_tool_use: true`.
128,000 tokens for the working context; the model reads far more total source text across its search iterations than fits in one window.
Runs an extended multi-query research process — response times are on the order of minutes, not the sub-10-second range of the search models. Not for interactive use.
$2 per 1M input and $8 per 1M output tokens, plus $2 per 1M citation tokens, $3 per 1M reasoning tokens and $5 per 1,000 search queries, and a report uses a lot of all four. Cost per report is noticeably higher and less predictable than the other Sonar models.
Ratings are estimated — limited independent data is available for this model.
Operator guidance
Reach for Deep Research only when you actually need an exhaustive, many-source written report and can wait minutes for it. For a cited answer to a hard question at interactive speed, Sonar Reasoning Pro is far cheaper and faster. For simple lookups, Sonar. Because Deep Research bills four usage axes (input, output, citation tokens, reasoning tokens) plus a per-query fee, budget it per-report, not per-token, and cap usage where the answer doesn't warrant the spend.
Use cases
- Comprehensive topic reports and literature reviews synthesised from many sources with citations
- Market, competitive, financial, technology, or health analyses where breadth of sourcing matters more than latency
- One-shot 'research this and write it up' tasks that would otherwise be a human afternoon of searching
Limitations
- Minutes-scale latency — unusable for interactive or real-time applications
- Four billed usage axes plus a per-search-query fee make per-call cost high and hard to predict; the most expensive Sonar variant in practice
- 128K working context despite reading many more source tokens across iterations — very long individual sources can still be truncated
- Underlying base model, exact per-report source/citation counts, and max output length are not published by Perplexity
- Legacy Sonar Chat Completions surface migrating to the Agent API (support until 2026-09-27, Perplexity changelog)
- Capability ratings are qualitative from Perplexity's positioning; no independent published benchmark for the API model