Perplexity
•
Sonar Reasoning Pro
•
Released 
January 2025

Perplexity
Sonar Reasoning Pro

Chain-of-thought reasoning combined with live web search and inline citations for source-verified analytical answers.

Modality:
Text
Image
model ID
perplexity/sonar-reasoning-pro

Output Speed *

N/A
tok/s

Intelligence Index *

11.8
/ 100

Context Window *

128000
tokens

Input price

12
Anytoken

Output price

48
Anytoken
Sonar Reasoning Pro: Search-Grounded Chain-of-Thought Reasoning with Citations Sonar Reasoning Pro is Perplexity's premier reasoning model, built on DeepSeek R1 with chain-of-thought and paired with real-time web search. It sits at the top of the Sonar reasoning tier, above the standard Sonar and Sonar Pro search models and below the exhaustive Sonar Deep Research model. Its defining trait is combining multi-step deliberation with live retrieval and inline citations in a single API call. That makes it strongest for analytical questions that require reconciling current web sources, rather than pure code generation or offline reasoning. Add Sonar Reasoning Pro to your stack via the AnyAPI.ai API.

Performance

Where Sonar Reasoning Pro Earns Its Place: Sourced Analytical Answers

Sonar Reasoning Pro is strongest on hard research questions that require reconciling conflicting sources and reasoning about recent developments. Independent testing places it well above average for a model in its tier: it scores 25 on the Artificial Analysis Intelligence Index, roughly double the average of comparable models. Because it merges chain-of-thought deliberation with live retrieval and inline citations, teams can trace each claim back to a source URL. The production consequence is a collapsed pipeline: retrieval, reasoning and citation happen in one API call rather than a hand-built search-plus-LLM chain.

Benchmarks

How Sonar Reasoning Pro Scores in Independent Testing

On the Artificial Analysis Intelligence Index, Sonar Reasoning Pro scores 25, placing it well above the average of about 14 for comparable models. That index is a composite covering reasoning, knowledge, mathematics and coding. Separately, third-party summaries report it ranking among the top models in Search Arena evaluations, statistically tied with Gemini-2.5-Pro-Grounding on search-grounded tasks. These results reflect its dual design: reasoning depth plus retrieval breadth. Provider-level latency and output-speed benchmarks are not consistently published for this specific endpoint, so treat throughput expectations cautiously and validate against your own workload.

Output Speed

*
N/A
tok/s

Intelligence Index

*
11.8
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE *

Adherence to multi-step structured instructions.
0
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Sonar Reasoning Pro accepts text input and returns text, with a 128K-token context window and up to 8,000 output tokens per response. That output cap is the most consequential limit: it includes the chain-of-thought <think> section, so verbose reasoning eats into the space available for the final cited answer. Outputs from reasoning models are structured as a <think> block followed by the answer, which downstream parsers must handle. Web search, citations, and search filters (domain and recency) are core capabilities, but structured-output and image-input behavior varies by endpoint and gateway.
Verified Specifications — 
Sonar Reasoning Pro
*
Input modalities
Text
Image
output modalities
Text
Context window
128000
 tokens
Maximum output tokens
115200
Reasoning
Yes
Knowledge cutoff
January 2025
Pricing (standard)
12
 AnyTokens in
 / 
48
 AnyTokens out

Comparison

Sonar Reasoning Pro vs Sonar Pro: Reasoning Depth or Context Reach?

Sonar Reasoning Pro and Sonar Pro are the two premium search models in Perplexity's Sonar family, and both add live retrieval and citations to a single API call. The practical decision is depth versus reach. Sonar Reasoning Pro layers DeepSeek R1 chain-of-thought on top of retrieval, producing structured, multi-step analysis, and scores markedly higher on the Artificial Analysis Intelligence Index (25 vs 9). Sonar Pro is a non-reasoning search model with a larger 200K-token context window versus 128K here, better suited to long documents and extended follow-up threads.

Dimension
Sonar Reasoning Pro
Sonar Pro
Context window *
128000
tokens
200000
tokens
Output speed *
N/A
tok/s
N/A
tok/s
Intelligence Index *
11.8
7.6
Input pricing
12
AnyToken
18
AnyToken
Output pricing
48
AnyToken
90
AnyToken
Knowledge cutoff *
January 2025
January 2025

Choose Sonar Reasoning Pro when the task demands step-by-step logical analysis over current sources — policy comparison, due diligence, or reconciling conflicting evidence — and you can tolerate slower, more deliberate responses. Choose Sonar Pro when you need broad, fast multi-source synthesis, larger context for long inputs, or lower-latency answers without a reasoning trace to parse. For simple factual lookups, both are overkill; drop to standard Sonar instead.

Limitations & Trade-offs

Where Sonar Reasoning Pro falls short

1
Small maximum output shared with reasoning. Output is capped at 8,000 tokens per response, and that budget includes the chain-of-thought <think> section. Verbose reasoning on a hard question can crowd out the final cited answer, forcing truncation or follow-up calls. Workloads that need long structured reports — full research briefs or multi-section documents — are affected most. For exhaustive report generation, Sonar Deep Research is the better fit.
2
Text-only input. Sonar Reasoning Pro processes text and returns text; image input with structured outputs is not supported in the thinking models, and modality behavior varies across endpoints. If your workload involves analyzing screenshots, charts, PDFs as images, or other visual inputs, this model cannot help and you need a multimodal alternative such as a vision-capable frontier model.
3
Search-grounded, not offline reasoning. The model's strength is analysis over live web results, so it always leans on retrieval and citations. For deterministic, closed-book tasks — reasoning purely over supplied context with no external search, or reproducible offline evaluation — the automatic search behavior and per-call search cost add overhead you may not want. A general reasoning model without forced retrieval is more predictable there.
4
Reasoning trace requires parsing and adds latency. Because reasoning models emit a <think> block before the answer, downstream systems must reliably separate reasoning from the final response, and structured-output requests with a new schema can take 10–30 seconds to prepare. Interactive, latency-sensitive chat surfaces may feel sluggish. If fast first-token response matters more than deliberation depth, a non-reasoning search model is preferable.

Best-Fit Workloads

Where this model earns its place

01

Source-verified research analysis

‍
The model's core fit: analytical questions that require reconciling conflicting web sources and reasoning about recent developments. It searches more sources than standard Sonar Reasoning, applies chain-of-thought deliberation, and returns structured answers with inline citations. Competitive intelligence, market analysis and technology scouting benefit directly. Note the 8K output cap limits how long any single report can be.

02

Policy and regulatory analysis

‍
Regulatory and policy questions require structured reasoning across evolving standards, guidelines and jurisdictional differences — exactly the retrieval-plus-reasoning pattern Sonar Reasoning Pro is built for. Its citations let compliance and legal-research teams trace each conclusion back to a source, which matters when answers inform decisions. Confirm currency of sources with the recency filter, since the model depends on live retrieval rather than static training data.

03

High-trust factual Q&A with attribution

‍
For product features where every answer must cite its evidence, Sonar Reasoning Pro returns URLs pointing to the web sources behind each response, collapsing a search-plus-LLM pipeline into one call. This suits knowledge assistants and analyst tools where accuracy and traceability outweigh raw speed. Use search_domain_filter to constrain answers to trusted domains and improve citation quality.

04

Complex multi-step decision support

‍
Tasks that build a structured argument from scattered web evidence — investment theses, vendor evaluations, feasibility assessments — benefit from the model's combination of retrieval breadth and reasoning depth. Its above-average Intelligence Index score reflects strength on this class of problem. Because responses include a reasoning trace, budget for parsing and for the deliberation latency that comes with chain-of-thought.

Pricing in anytokens via AnyAPI
Input
12
₳
Output
48
₳
Cache write
—
₳
Cache read
—
₳

Integration

Access Sonar Reasoning Pro via AnyAPI.ai

Access Sonar Reasoning Pro through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Sonar Reasoning Pro without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Sonar Reasoning Pro and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Sonar Reasoning Pro against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Sonar Reasoning Pro from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Sonar Reasoning Pro for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Sonar Reasoning Pro is best for analytical research questions that require reasoning over current web sources with citations. Built on DeepSeek R1 chain-of-thought and paired with live search, it excels at reconciling conflicting sources, policy and regulatory analysis, and high-trust factual Q&A where every claim must trace back to a source URL.

Sonar Reasoning Pro has a 128,000-token context window and supports up to 8,000 output tokens per response. Importantly, that output limit includes the chain-of-thought reasoning section, so long deliberation reduces the space available for the final cited answer. For very long reports, Sonar Deep Research is more appropriate.

No. Sonar Reasoning Pro accepts text input and returns text; image input with structured outputs is not supported in the thinking models. It does not offer conventional function/tool calling. Instead, it provides built-in web search with domain and recency filters, plus JSON-schema structured outputs on Perplexity's API, though behavior varies by endpoint.

Sonar Reasoning Pro adds DeepSeek R1 chain-of-thought reasoning on top of search and scores much higher on the Artificial Analysis Intelligence Index (25 vs 9). Sonar Pro is a non-reasoning search model with a larger 200K-token context window. Choose reasoning for analytical depth; choose Sonar Pro for broader context and faster answers.

Sonar Reasoning Pro is available through Perplexity's OpenAI-compatible API using the model ID sonar-reasoning-pro, and through gateways such as AnyAPI.ai. Send a standard chat-completion request; responses include a <think> reasoning section followed by a cited answer, which you should parse separately in production.

* Benchmark data source: Artificial Analysis artificialanalysis.ai