Perplexity challenges Google and OpenAI with Sonar search API

If you have ever built an app around raw LLM APIs, you know the frustration of the static knowledge cutoff. Unless you want to spend weeks stitching together web scrapers, embeddings, and complex retrieval pipelines, your app is essentially operating with historical data.

Now, fresh off a $500M funding round that pushed its valuation to $9 billion, Perplexity wants to take care of the plumbing. The startup just launched its Sonar API, directly offering real-time web search with automatic source citations to developers and enterprises.

Fast lookups versus deep research

The API comes in two flavors: base Sonar and Sonar Pro.

Base Sonar is built for speed and budget efficiency. It handles context windows up to 127,000 tokens and is aimed at quick, responsive queries where a single search pass is enough.

For heavy lifting, Sonar Pro expands the context window to 200,000 tokens. Instead of a single query, Pro runs multiple searches in the background to tackle complex prompts, returning twice as many source citations as the standard version so downstream apps can verify facts.

In OpenAI's SimpleQA benchmark, Sonar Pro posted an F-score of 0.858, topping base Sonar’s 0.773. Perplexity claims it outperformed Google's grounded search models, though publications like The Decoder pointed out that comparing it to stock GPT-4o or Claude is slightly lopsided since those models lacked live web access in the run.

Perplexity has not shared full architectural details for the underlying models, but previously acknowledged that an earlier iteration in the Sonar family was a fine-tuned version of Meta's open-source Llama 3.1 70B.

The pricing formula and early adopters

The pricing blends fixed search fees with standard token usage.

Both tiers charge a flat $5 per 1,000 web searches. Beyond that, base Sonar charges $1 per 750,000 words across input and output. Sonar Pro charges $3 per 750,000 input words and $15 per 750,000 generated words.

Developers can constrain the exact domains Sonar draws from, and Perplexity emphasized that enterprise prompt data will never be used to train its models.

Big names are already plugging it in. Zoom integrated Sonar into AI Companion 2.0 to give meeting participants live web lookups without leaving their video feed, while sales platform Copy.ai claims it saves customers hours of weekly manual lookup.

Why it matters

We think this is Perplexity’s most pragmatic strategic move to date. Raising half a billion dollars at a $9B valuation means consumer subscriptions alone will not cut it; the company needs durable enterprise revenue.

For builders, Sonar solves an annoying problem. Building production-grade RAG over the open web requires orchestrating search providers, parsing noisy HTML, and hoping the LLM attributes sources correctly. Packaging all of that behind a single endpoint makes it instantly appealing for workflow tools, automated research agents, and customer support bots. If you need verifiable, up-to-date data in your product, it is well worth testing now.

The real test will be reliability. Real-time web retrieval is inherently non-deterministic, and Perplexity still faces scrutiny over verbatim scraping and the occasional hallucinated link. Moreover, OpenAI and Google already own the underlying search or frontier model layers. Perplexity has established an early lead in packaged search RAG, but keeping that edge will get much harder as hyperscalers roll out integrated web grounding at commodity rates.